Work · Sources

AI at work

Also called ChatGPT work · Copilot at work · workplace AI · generative AI job · paste into chatbot · AI draft · AI replace my job · confidential ChatGPT · AI email · job automation

Named sources. May contain inaccuracies or be incomplete. Not medical, legal, financial, or other professional advice.

AI can help with a work draft, but useful assistance depends on the task, the information you share and the check you can perform. Start with one permitted task using public or approved material. Judge the finished work, including review and correction, while keeping responsibility for the result.

What a workplace AI assistant does

A draft, a verified answer and an action taken on your behalf require different kinds of trust.

Give the assistant a defined piece of work

Generative AI creates content such as text or code from instructions and supplied material. A large language model is one kind of generative system: it learns patterns in language and uses them to produce a response. Your prompt is the request and context you give it. That can help turn a rough idea into an outline, rewrite a paragraph or propose questions. It does not give the system automatic access to your company's actual promises, policies or records. Brynjolfsson, Li and Raymond distinguish the underlying model from the workplace system built around it. Begin by naming the output you want and which facts it is allowed to use.

Generative AI at Work — Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 140(2), 889–942 (2025)ChatGPT and large language models: what's the risk? — UK National Cyber Security Centre (2023)

Use source material to make an answer checkable

Some assistants can search documents or retrieve passages before generating an answer. Retrieval means finding material; generation means composing the response. These are separate operations: finding the right document does not prove that the summary preserves its meaning. The support assistant studied by Brynjolfsson and colleagues provided suggestions and documentation links, while agents retained discretion over the conversation. At your desk, a useful answer includes a passage or locator you can open, not merely a convincing reference to a handbook. Check that the cited passage supports the particular statement and that the document is the current one.

Generative AI at Work — Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 140(2), 889–942 (2025)AI RMF Playbook — MEASURE, NIST

Review an action before granting permission to carry it out

An assistant that drafts an email differs from an agent that can send it. Agentic AI is connected to tools and can take actions toward a goal, potentially without continuous human intervention. The NCSC's May 2026 guidance says organizations should define ownership, access, monitoring and the ability to stop the system before connecting agents to real work. For an ordinary user, the practical distinction is straightforward: a draft can be edited; a sent message, changed record or executed command may already affect someone. Use only an approved arrangement, with the access and review your organization requires. A successful drafting trial does not establish readiness for autonomous action.

Thinking carefully before adopting agentic AI — UK National Cyber Security Centre (May 15, 2026)

Where it can fit into an ordinary workday

These are practical applications of the evidence, not tested STUDlearn procedures or promised savings.

Start an email with its purpose and audience

When you know what an email must accomplish but struggle to start, ask for a structure using public facts or fictional placeholders. For example: draft a brief meeting invitation with spaces for the date, location and reply deadline. Noy and Zhang tested occupation-specific writing assignments and found a useful role for assistance with composition. Their tasks did not depend on private organizational knowledge or precise factual accuracy, so the result cannot certify a real client message. Put the correct facts into the draft on an authorized system. Read it for tone, invented commitments and whether the recipient can tell what to do next. You remain responsible for the send.

Experimental evidence on the productivity effects of generative artificial intelligence — Noy and Zhang, Science (2023), author-hosted manuscript The productivity effects of generative artificial intelligence — Noy and Zhang, VoxEU (2023)

Compare a summary with the document it represents

When a long notice or meeting record feels hard to digest, an approved assistant may help organize it. A summary is a shorter account of a source, so its quality depends on preserving decisions, exceptions and uncertainty rather than merely sounding clear. NIST recommends evaluating AI in the context in which it will be used. For a routine summary, compare the output with the original and ask: did it miss a decision, turn a suggestion into a commitment, or attach the wrong person or date? Keep the record available for the reader. If you cannot open the underlying source or competently review it, the summary is not a reliable basis for your decision.

Consider a fictional notice: a workshop is planned for Thursday, subject to the room being confirmed, and attendees must reply by Tuesday. An assistant summarizes it as 'The workshop is Thursday; reply by Thursday.' The sentence is fluent, but it has removed a condition and changed a deadline. A source check has to compare those details, not merely reread the summary for good grammar.

The reviewer restores 'planned, pending room confirmation,' corrects the reply date and leaves a source locator. If the room is later confirmed, that is a new fact requiring an updated record. Counting the quick first draft as time saved would omit this review and correction work. This invented example shows what responsible review actually checks; it does not establish a particular tool's accuracy or authorize sharing a real internal notice with it.

AI RMF Playbook — MEASURE, NIST Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST AI 600-1 (2024)

Check the condition and the deadline against the source

All three excerpts below concern the same fictional workshop notice.

Source notice · section 1

The workshop is planned for Thursday, subject to the room being confirmed. Attendees must reply by Tuesday.

Draft being checked

The workshop is Thursday; reply by Thursday.

Two material facts to restore
Detail	What the source actually says
Event status	Planned; room confirmation is still required.
Reply deadline	Tuesday, before the planned Thursday event.

After source comparison

The workshop is planned for Thursday, pending room confirmation. Reply by Tuesday.

Source: fictional workshop notice, section 1.

The revision restores the condition and deadline. It does not confirm the room. That would require a new fact from the responsible source.

This short specimen is complete for the two facts shown. One checked excerpt cannot certify a whole document or authorize entering real workplace information into a tool.

Choose a permitted route before supplying work information

When the task involves a client file, staff record or unreleased number, decide whether the tool and that particular use are authorized before entering anything. Confidentiality concerns what may be shared, with whom and for what purpose; it is not just a setting labeled private. The NCSC describes shadow AI as use outside an organization's approved systems and processes, which can reduce visibility and control of information. Replace actual details with invented examples only when they no longer reveal the underlying matter. Otherwise use the approved environment or do the work without the tool. An app installed on a work laptop is not, by itself, permission to upload every file.

The hidden risks of shadow AI — UK National Cyber Security Centre (September 7, 2026)How should we assess security and data minimisation in AI? — UK Information Commissioner's Office Formal Opinion 512: Generative Artificial Intelligence Tools — American Bar Association (July 29, 2024)

Discuss a changing task rather than your whole job at once

When a headline makes your occupation sound obsolete, identify one task the tool may change and one responsibility that remains. Drafting a report, verifying its figures and explaining its recommendation are different parts of a job. The ILO's 2025 index measures potential exposure of occupational tasks to generative AI; it does not count layoffs. Autor's task framework explains why automation can substitute for some work and complement other work. This does not guarantee employment security. It makes a workplace conversation more useful: what will be automated, what expertise is needed to review it, and what training, staffing or sign-off will support the new process?

Generative AI and Jobs: A Refined Global Index of Occupational Exposure — ILO Working Paper 140 (2025)Why Are There Still So Many Jobs? The History and Future of Workplace Automation — David Autor, Journal of Economic Perspectives 29(3), 3–30 (2015)

What the workplace studies actually found

The studies answer different questions. Their results describe particular people, tasks and systems, rather than a single productivity rate for AI.

Some short writing tasks improved with AI assistance

Shakked Noy and Whitney Zhang randomly gave ChatGPT access to half of 453 college-educated professionals completing paid, occupation-specific writing tasks in 2023. Participants were managers, marketers, grant writers, consultants, human-resource professionals and data analysts. Average completion time fell about 40 percent and judged quality rose about 18 percent. Lower-baseline writers gained more. The assignments included reports, press releases and tactful emails, and experienced graders assessed the output. The authors emphasize that the tasks favored clear, relatively generic writing and did not require context-specific knowledge or precise factual accuracy. That makes the finding useful for drafting, while leaving a real workplace's fact checking, authorization and audience requirements to be tested separately.

Experimental evidence on the productivity effects of generative artificial intelligence — Noy and Zhang, Science (2023), author-hosted manuscript

A useful draft can change how people work without proving long-term competence

In their public account, Noy and Zhang reported that many participants submitted unedited or lightly edited chatbot output and graders often rated it well. That describes what worked in their experiment; it is not an instruction to forward raw output at work. Participants enjoyed the assisted task more, and both excitement and concern about AI rose temporarily. Those attitude effects faded in follow-up surveys, while treated participants were more likely to report later workplace use. Adoption, enjoyment and a good graded document are different outcomes from sustained competence or a client's satisfaction. Keep that distinction when someone offers a compelling demonstration as proof of lasting value.

The productivity effects of generative artificial intelligence — Noy and Zhang, VoxEU (2023)Experimental evidence on the productivity effects of generative artificial intelligence — Noy and Zhang, Science (2023), author-hosted manuscript

Support workers gained unevenly from a specialized suggestion tool

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied staggered access to a conversational assistant among 5,172 customer-support agents. The published 2025 study found about 15 percent more issues resolved per hour on average. Less experienced and lower-skilled agents improved speed and quality; the most experienced and highest-skilled agents had small speed gains and small quality declines. Customers were more polite and less likely to request a manager. Agents could edit or ignore suggestions. This was a tool adapted to one firm's technical-support work, rather than an unrestricted public chatbot. A team should therefore examine whose work improves and whether quality changes, rather than assign the average gain to everybody.

Generative AI at Work — Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 140(2), 889–942 (2025)

Check capability on the task you actually need done

Dell'Acqua and colleagues call this the jagged technological frontier: AI can help with one task and hurt performance on another that looks equally difficult to a person. Their experiment included 758 BCG consultants, roughly 7 percent of individual contributors, in early stages of their consulting careers. Participants received no AI, GPT-4, or GPT-4 with a prompting overview. Across 18 tasks within the tested capability range, AI users completed 12.2 percent more tasks, worked 25.1 percent faster and produced higher-rated work. Lower-baseline performers gained more. On a deliberately selected task outside that range, the combined AI groups were about 19 percentage points less likely to reach the correct answer. The 2026 journal article preserves the key contrast. It does not locate today's boundary for your task or certify your preferred prompt.

Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua and colleagues, Organization Science (2026)

Measure completed work rather than the feeling of moving faster

METR randomized AI access on 246 real tasks completed by 16 experienced open-source developers in familiar repositories during early 2025. AI access made completion take about 19 percent longer, although developers believed it had helped them work faster. This is a small, specific study, not proof that coding assistance always slows work. In February 2026, METR reported that a later experiment's results were unreliable as a measure of current uplift because participation, task selection and concurrent-agent time measurement had changed. The researchers thought gains were likely larger by then, but could not reliably quantify the increase. Both reports support checking your own finished workflow rather than promoting an old result into a timeless claim.

Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METR Update on AI developer productivity — METR (February 24, 2026)

Exposure estimates describe potential change rather than observed job loss

The ILO–NASK 2025 index combines worker task assessments, expert input and model predictions to estimate how generative AI could affect occupations. About one in four workers globally are in occupations with some exposure, and 3.3 percent of employment falls in the highest category. Clerical work remains especially exposed; exposure also differs by income group and sex because occupational employment differs. The authors consider transformation more likely than complete replacement because many jobs retain tasks requiring human input. These are estimates of potential, not a count of displaced people, a forecast for a particular employer, or reassurance that no job will disappear. Use them to identify where transition support may be needed.

Generative AI and Jobs: A Refined Global Index of Occupational Exposure — ILO Working Paper 140 (2025)One in four jobs at risk of being transformed by GenAI, new ILO–NASK Global Index shows — ILO

Handle work information before it enters a tool

Check approval for the task, data and system together

A company may approve one AI service for one class of information while restricting another use. Before uploading, establish what data the approved account may receive, who can access inputs and outputs, how they are retained, and whether connected services are covered. The NCSC's September 2026 shadow-AI guidance explains how unapproved use can bypass established visibility and controls. If the policy is unclear, bring a concrete example of the proposed task to the policy owner without exposing the actual confidential file. Approval answers whether you may use the tool; it does not answer whether its output is correct.

The hidden risks of shadow AI — UK National Cyber Security Centre (September 7, 2026)ChatGPT and large language models: what's the risk? — UK National Cyber Security Centre (2023)

Use only the information needed for the purpose

Data minimisation means using the personal information necessary for a defined purpose, rather than providing everything that might be convenient. The UK's ICO treats this as a continuing requirement when organizations use AI. At a desk, an editorial application is to replace a real person's details with a genuinely fictional example when you only need help with wording or structure. Removing a name alone may leave a person identifiable through role, dates or circumstances. Keep patient and student records, client matters, employee details, privileged material, credentials, restricted code and unreleased figures out of unapproved tools. A full-file upload deserves more scrutiny than a public paragraph. This is privacy literacy, not a jurisdiction-specific compliance determination.

How should we assess security and data minimisation in AI? — UK Information Commissioner's Office Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST AI 600-1 (2024)The hidden risks of shadow AI — UK National Cyber Security Centre (September 7, 2026)

Treat confidentiality as a duty attached to the information

ABA Formal Opinion 512 explains lawyers' duties of competence, confidentiality, communication, supervision and reasonable fees when using generative AI. Its discussion of self-learning tools says informed consent can be necessary before entering information relating to a representation, and generic boilerplate is insufficient. That opinion concerns lawyers and US model ethics rules; it does not impose one consent procedure on every profession. Teachers, clinicians, accountants and government staff must use their own applicable rules and approved processes. The common workplace question is whether this transfer is allowed before it happens. Deleting a conversation later cannot retroactively authorize a disclosure or restore control of a copy already transferred.

Formal Opinion 512: Generative Artificial Intelligence Tools — American Bar Association (July 29, 2024)ChatGPT and large language models: what's the risk? — UK National Cyber Security Centre (2023)

Separate a provider's privacy promise from your permission to share

A promise not to train on customer data addresses one use of information. It does not, on its own, answer retention, access, security or your employer's authorization requirements. FTC staff explain that providers can face liability when they break privacy and confidentiality commitments, including promises about model training. That is guidance directed at providers, rather than a declaration that a particular work file is safe to upload. Likewise, a prompt does not necessarily update a model instantly. Read the current approved terms and workplace policy for the system you actually use. Treat the transfer of restricted information as the decision to resolve before composing a prompt.

AI Companies: Uphold Your Privacy and Confidentiality Commitments — FTC Office of Technology (January 2024)ChatGPT and large language models: what's the risk? — UK National Cyber Security Centre (2023)

What a useful check looks like

Compare claims with evidence outside the conversation

NIST calls confidently false generated content confabulation, often called hallucination. Automation bias means giving an automated suggestion more weight than its evidence deserves. These concepts explain why a polished answer can pass a casual read while still containing the wrong figure or unsupported conclusion. For a factual draft, compare names, dates, amounts and citations with the original record. A citation should exist, address the question and support the sentence. For a calculation, use the authoritative data and an appropriate independent calculation. Asking the same assistant whether it is sure can identify a question to investigate, but its assurance is not that investigation.

Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST AI 600-1 (2024)AI RMF Playbook — MEASURE, NIST Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua and colleagues, Organization Science (2026)

Give the reviewer evidence and authority to change the result

Human review has value when someone has the knowledge, time and source access to identify an error and can reject or correct the output. In a 2026 MIT account of research on 72 BCG consultants, researchers observed GPT-4 defending recommendations with increasingly persuasive language when users challenged it. That analysis describes interactions in one task, not intentional deceit or every model's behavior. Its practical relevance is that remaining inside a persuasive chat may be a weak check. Move to the original records and record the unresolved point. A named reviewer who merely clicks approval without that comparison has not established reliability.

How generative AI 'persuasion bombs' users — and how to fight back — MIT Sloan (April 28, 2026)AI RMF Playbook — GOVERN, NIST

Match review effort to the consequence of an error

NIST frames risk in terms of both the chance of an event and the severity of its consequences. A wrong heading in a private outline and a wrong number in a client decision have different stakes. Before reviewing, describe the specific failure: an invented deadline, missing exception, wrong recipient or unsupported recommendation. Choose a check capable of finding it, and involve the accountable specialist when the matter needs expertise you lack. Safety-critical, licensed or regulated work retains its required competence and sign-off. A generic 'human checked' label cannot replace that process. If an essential claim remains unverified, keep the output as a draft or use the existing authorized workflow.

Artificial Intelligence Risk Management Framework (AI RMF 1.0) — NIST AI 100-1 (2023)Formal Opinion 512: Generative Artificial Intelligence Tools — American Bar Association (July 29, 2024)AI RMF Playbook — MANAGE, NIST

Test the complete workflow against a defined standard

Choose the success criteria before a permitted comparison. For a meeting summary, that might mean retaining all decisions, owners and deadlines, preserving uncertainty, and adding no unsupported commitments. Count time spent preparing the input, generating, reviewing and correcting, not just time until text appears. Compare tasks of similar difficulty against the same quality standard. Danielle Li's public guidance stresses defining success; MIT's 2026 workplace synthesis adds evidence before scaling and attention to rework. This small comparison is a practical adaptation, not a controlled trial. One good outcome supports another cautious test of that task, rather than an all-job productivity percentage.

How to use generative AI to augment your workforce — MIT Sloan (February 24, 2025)10 levers shaping whether generative AI truly improves worker performance — MIT Sloan (September 3, 2026)AI RMF Playbook — MEASURE, NIST Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METR

Give the task an owner, a purpose, a check and a fallback

NIST's voluntary AI Risk Management Framework organizes risk work into four connected functions: GOVERN sets responsibilities and rules; MAP describes the use, affected people and possible harms; MEASURE evaluates performance and risk; MANAGE responds to what is found. They are not an ordered checklist or a certificate of safety. Its trustworthiness characteristics cover validity and reliability, safety, security and resilience, accountability and transparency, explainability and interpretability, privacy, and fairness with harmful bias managed. A desk-level application is to write who owns the result, what the assistant may do, which failure your check can catch and what happens if it fails. The generative-AI profile adds detail on confabulation, privacy, information security, intellectual property and human over-reliance. Correct wording is only one part of responsible use.

Artificial Intelligence Risk Management Framework (AI RMF 1.0) — NIST AI 100-1 (2023)AI RMF Core — NIST AI Resource Center Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST AI 600-1 (2024)

Where workplace decisions remain unsettled

Ask what training and consultation will change in practice

OECD surveys of 5,334 workers and 2,053 firms in manufacturing and finance across seven countries found generally positive reported effects alongside concern about job loss. Training and worker consultation were associated with better reported outcomes. Association means the factors occurred together; it does not establish that training caused the improvement. Organizations with good management may differ in other ways. The 2023 report concerns surveys conducted in early 2022, before ChatGPT's public release; it is not a survey of today's generative tools. The genuine unresolved question is which training and consultation arrangements work best for which jobs. Ask who can demonstrate a useful review, who workers can report problems to, and whether staff feedback changes deployment. Optimism and worry both deserve a specific account of what happened.

The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers — Lane, Williams and Broecke (2023)OECD AI surveys of employers and workers — project overview Kate Kellogg — MIT Sloan faculty directory

Match assistance and review to the worker's experience

The writing, support and consultant studies often found larger gains for lower-baseline workers within the tasks where assistance worked. That is a shared pattern, rather than opposing findings about novices. Yet the consultant study also demonstrated a correctness drop outside the tested capability range. 'Lower skill' there was relative to other highly trained consultants, not an entry-level worker in every field. A team can reasonably debate which tasks need expert review, how much training new staff require and whether experts should use different workflows. Narrowing an output-quality gap does not prove that professional knowledge, exception handling or supervisory competence has become unnecessary.

Experimental evidence on the productivity effects of generative artificial intelligence — Noy and Zhang, Science (2023), author-hosted manuscript Generative AI at Work — Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 140(2), 889–942 (2025)Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua and colleagues, Organization Science (2026)

Use the disclosure rules that govern the work

Whether AI assistance must be named depends on the employer, client agreement, professional requirements and circumstances. ABA Opinion 512 describes situations requiring communication or informed consent and situations where routine assistance need not be disclosed; it is not a universal email-label rule. A useful internal handoff can say what was checked, who reviewed it and what remains uncertain, while following any required AI disclosure. Separately, disclosure of a tool used to evaluate a worker is a different issue from disclosure that you used it to draft a paragraph. Ifeoma Ajunwa's workplace-technology scholarship helps keep that distinction visible. Transparency should help the affected person understand the decision and the route for questioning it.

Formal Opinion 512: Generative Artificial Intelligence Tools — American Bar Association (July 29, 2024)Ifeoma Ajunwa — Emory Law faculty profile AI RMF Playbook — GOVERN, NIST

Read employment forecasts alongside the choices that shape jobs

Task exposure, employer adoption and observed employment are different measurements. A forecast should state its date, geography, occupations and assumptions about demand and deployment. Autor explains that technology can substitute for labor while also creating complementary work and changing demand. This does not rule out displacement. In their April 2026 synthesis of early experiments at more than 20 companies, Ben Armstrong and Julie Shah describe substantial variation in how organizations turn AI into supervisory work and whether that work is engaging or repetitive. It is qualitative organizational evidence, not a causal estimate of net jobs. Training, staffing, workflow design and how saved time is used remain real decisions.

Why Are There Still So Many Jobs? The History and Future of Workplace Automation — David Autor, Journal of Economic Perspectives 29(3), 3–30 (2015)Generative AI and Jobs: A Refined Global Index of Occupational Exposure — ILO Working Paper 140 (2025)Humans in the Loop: The Evolution of Work in Early Experiments with Generative AI — Armstrong and Shah, MIT Industrial Performance Center (April 2026)

Keep enough skill to supervise the work

Separate assisted performance from what you can do later

Hamsa Bastani and colleagues tested two GPT-4-based tutors in high-school mathematics: an open-ended helper and a tutor with teacher-informed safeguards. Both improved assisted practice performance, but the open-ended group later scored worse than controls on unaided exams. The safeguarded tutor largely prevented that penalty without demonstrating a later learning gain. The published 2025 study is classroom evidence, not a trial of adult workplace use. It makes a useful question visible: did the tool help you learn a method or simply produce an answer? When the job still requires judgment, close the chat and explain the calculation, exception or recommendation using the source material. Use Learning and remembering for the fuller study-skills question.

Generative AI without guardrails can harm learning: Evidence from high school mathematics — Bastani and colleagues, PNAS 122(26), e2422633122 (2025)

Use your knowledge to direct and evaluate the assistant

A 2025 CHI survey by Hao-Ping Lee, Advait Sarkar, Lev Tankelevitch and colleagues collected 936 examples from 319 knowledge workers. Higher confidence in generative AI was associated with less reported critical-thinking effort, while greater confidence in one's own ability was associated with more. The researchers described a shift toward verification, integrating responses and overseeing tasks. This was self-reported behavior, not an objective test proving that AI destroys thinking skills. The support-agent study also found evidence consistent with learning when assistance was temporarily unavailable. Outcomes can differ with design and use. Maintain chances to practice, explain and challenge the work you remain accountable for, rather than treating either universal learning or universal deskilling as settled.

The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — Lee, Sarkar, Tankelevitch and colleagues, CHI (2025)Generative AI at Work — Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 140(2), 889–942 (2025)

Keep the process useful over time

Keep the current policy and an accountable contact within reach

Save the current workplace AI policy or the name of the person responsible for the approved service. Check whether the permission covers your specific task and class of information when either changes. NIST's GOVERN guidance connects AI responsibilities to existing organizational rules, while the NCSC explains the risks when informal use escapes those processes. An unclear policy is a reason to seek a concrete answer before supplying restricted material. It does not require abandoning a useful public-material experiment that is already permitted. Policy gives you a boundary; source comparison gives you evidence about the result.

AI RMF Playbook — GOVERN, NIST The hidden risks of shadow AI — UK National Cyber Security Centre (September 7, 2026)

Keep sensitive facts on the system authorized to hold them

Write a reusable prompt around the task's structure rather than real client names, personal histories or unreleased numbers. Fill those facts into the result only in an authorized place, after checking that the wording still fits. ICO minimisation guidance provides the principle of limiting personal data to what the purpose requires. The desk-level example is an editorial adaptation: it is useful when generic structure is enough, but unsuitable if the fictionalized input still reveals a person or the real task cannot be reviewed without restricted information. In that case, use the approved data-handling route or the ordinary workflow.

How should we assess security and data minimisation in AI? — UK Information Commissioner's Office Formal Opinion 512: Generative Artificial Intelligence Tools — American Bar Association (July 29, 2024)

Keep a source locator with factual work

Store the approved source location, date or document version beside a factual draft so someone can repeat the comparison later. A sentence such as 'the notice dated October 1, section 3' is useful only if it identifies a real record the reviewer can access. Where records conflict, state what differs and seek the responsible person's resolution; a fluent generated paragraph should not silently choose. NIST's measurement guidance calls for context and documented evaluation. The ICO also distinguishes a model's average predictive accuracy from accurate information about a particular person. If a record contains a prediction about someone, identify it as an inference rather than an established fact. Neither a general accuracy score nor a tidy citation resolves a specific contradictory record.

AI RMF Playbook — MEASURE, NIST What do we need to know about accuracy and statistical accuracy? — UK Information Commissioner's Office

Be able to explain the decision you approve

Before approving work that requires your expertise, explain its reasoning with the authoritative source open and the chatbot closed. Identify the evidence, the exception and what would change your conclusion. This is a practical accountability check, rather than an exam score or a validated anti-deskilling intervention. Bastani's study shows why assisted performance and unaided understanding need separate attention, while the CHI survey suggests checking and task oversight may become a larger part of knowledge work. If you cannot reconstruct a material decision, obtain competent review or do that part yourself before relying on the output.

Generative AI without guardrails can harm learning: Evidence from high school mathematics — Bastani and colleagues, PNAS 122(26), e2422633122 (2025)The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — Lee, Sarkar, Tankelevitch and colleagues, CHI (2025)AI RMF Core — NIST AI Resource Center

Keep a small permitted record and revisit recurring misses

Where workplace rules allow it, note the task, approved tool, check performed and one useful correction. Do not create a private archive of confidential prompts or copy restricted files into a personal log. NIST's MANAGE guidance supports monitoring problems and responding to them, and Kellogg's implementation work emphasizes workers' involvement and training. Review recurring misses when the task, source or tool changes. If assistance repeatedly invents commitments or consumes more correction time than it saves, narrow its role or use the established workflow. When AI influences scheduling, hiring or performance assessment, ask what information went in, which output affected the decision, who is responsible for the final decision and how an error can be raised. These apply NIST's feedback principle; employer policy, worker representation and local rules determine the formal route, rather than a uniform legal right assumed here.

AI RMF Playbook — MANAGE, NIST Kate Kellogg — MIT Sloan faculty directory Ifeoma Ajunwa — Emory Law faculty profile AI RMF Playbook — GOVERN, NIST

Check the claims people repeat

An established finding stays bounded to what its study measured. Read the claim and its evidence together before extending it to your work.

Some writing gains are established within a defined experiment

Claim: ChatGPT-style tools can change time and quality on some mid-level writing tasks in a bounded experiment. Verdict: established. Noy and Zhang measured shorter completion times and higher grades. This does not establish a client-ready result or a guaranteed saving for your own job.

Experimental evidence on the productivity effects of generative artificial intelligence — Noy and Zhang, Science (2023), author-hosted manuscript

Task exposure leaves an occupation's future open

Claim: Generative AI will replace every job this year. Verdict: not. The ILO's index measures potential task exposure. It does not predict universal replacement, count layoffs or settle demand and adoption. A forecast and an observed employment change require their own evidence.

Resolve permission before sharing confidential material

Claim: Confidential client or employer material is safe in a public chatbot if I delete the chat. Verdict: not. Deletion does not retroactively authorize the transfer. Applicable confidentiality duties, approved terms and workplace rules govern the information before it enters the tool.

Formal Opinion 512: Generative Artificial Intelligence Tools — American Bar Association (July 29, 2024)AI Companies: Uphold Your Privacy and Confidentiality Commitments — FTC Office of Technology (January 2024)The hidden risks of shadow AI — UK National Cyber Security Centre (September 7, 2026)

Fluent wording still needs a factual check

Claim: If the chatbot wrote it fluently, it is ready to send. Verdict: not. Fluent false content is a recognized risk. Review against the original evidence and the task's required quality standard before relying on it.

Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST AI 600-1 (2024)Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua and colleagues, Organization Science (2026)

Evaluate gains for the particular worker and task

Claim: A workplace AI assistant raises every worker's performance equally. Verdict: not. The support study found different effects by experience and skill, and the consultant study found different effects by task. An average is not an individual guarantee.

Generative AI at Work — Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 140(2), 889–942 (2025)Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua and colleagues, Organization Science (2026)

Treat the training association as evidence with a causal limit

Claim: Training and consulting workers about AI is linked with better reported outcomes. Verdict: established for the OECD survey association. The stronger claim that training causes improvement everywhere remains unproven and open to debate. Association and universal effectiveness are different propositions.

The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers — Lane, Williams and Broecke (2023)

Keep the knowledge needed to explain and supervise your work

Claim: Using AI at work means I no longer need to know the material. Verdict: talk. Producing a draft does not demonstrate the competence needed to check, explain or approve it. The workplace still determines which responsibilities and expertise your role requires.

The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — Lee, Sarkar, Tankelevitch and colleagues, CHI (2025)Generative AI without guardrails can harm learning: Evidence from high school mathematics — Bastani and colleagues, PNAS 122(26), e2422633122 (2025)AI RMF Core — NIST AI Resource Center

010

Try one small, permitted task

Keep the first move small enough that you can judge it yourself.

Name the task, your check and the human step that remains

Choose one low-stakes task you are already allowed to do with AI, such as structuring a meeting invitation from public or fictional information. Write three lines: the assistant will propose the structure; I will check it against the request; I will supply the real facts, choose the tone and decide whether to send. These lines define the assistant's job and the decisions that remain yours. If the actual task needs restricted material, a consequential decision or expertise you lack, use the existing approved process instead of widening this experiment.

Why Are There Still So Many Jobs? The History and Future of Workplace Automation — David Autor, Journal of Economic Perspectives 29(3), 3–30 (2015)AI RMF Core — NIST AI Resource Center Thinking carefully before adopting agentic AI — UK National Cyber Security Centre (May 15, 2026)

Try a structure with genuine placeholders

Ask for a short outline with slots for Date, Location and Reply deadline, without entering actual restricted facts. Specify the audience and desired outcome, then compare the proposed structure with the original request. Keep wording that helps, remove any invented promise, and fill real details only on the system authorized to hold them. The writing experiment supports exploring composition help, while minimisation guidance supports keeping unnecessary personal data out. This exact exercise is an editorial example, not the researchers' protocol.

The productivity effects of generative artificial intelligence — Noy and Zhang, VoxEU (2023)How should we assess security and data minimisation in AI? — UK Information Commissioner's Office

Try one source comparison before relying on the result

Select one generated claim, number or commitment and open the record it should reflect. Mark it matches, wrong or not in the source. Correct or remove a miss, and repeat the review for every material fact before an actual send. This first comparison teaches the method; checking one sentence does not certify a whole document. Record how much review and correction the task needed, without saving secrets. A useful result is a finished piece of work you can explain and stand behind, including the checking time.

AI RMF Playbook — MEASURE, NIST AI RMF Playbook — MANAGE, NIST How to use generative AI to augment your workforce — MIT Sloan (February 24, 2025)

011

Use this with the rules for your work

Use workplace guidance and qualified sign-off for consequential work

This is a reading companion about ordinary workplace drafting, checking, privacy, over-trust, and dividing a job into tasks. It is not legal, human-resources, medical, or financial advice, a productivity promise, or a vendor recommendation. Use your employer’s current policy and approved tools when they exist. Keep client, patient, student, employee, and employer-confidential material out of tools you do not control. Safety-critical, licensed, or regulated work still needs the competence and sign-off the role requires. For studying with a chatbot, see Learning and remembering. For checking an online claim or an AI image, see Making sense of online information.

AI RMF Core — NIST AI Resource Center Formal Opinion 512: Generative Artificial Intelligence Tools — American Bar Association (July 29, 2024)The hidden risks of shadow AI — UK National Cyber Security Centre (September 7, 2026)

Read named findings as a companion to your judgment

Named sources. Honest paraphrase of the finding. Not medical, legal, or financial advice. Career and work covers the wider career and work-design question; this topic concentrates on using a workplace assistant. Learning and remembering covers study and memory. Making sense of online information covers evaluating online claims and AI images. Those topics complement this one without replacing the policy, evidence or competence your actual work requires.

AI RMF Core — NIST AI Resource Center Why Are There Still So Many Jobs? The History and Future of Workplace Automation — David Autor, Journal of Economic Perspectives 29(3), 3–30 (2015)

References for this topic

Sources and context

Generative AI at Work — Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 140(2), 889–942 (2025) ↗
id
ai-at-work--support-qje-2025 — Generative AI at Work — Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 140(2), 889–942 (2025)
kind
study
accessed at
2026-10-01
supports
Opened published article including abstract, setting, results and learning analysis. Verified 5,172 agents, staggered access, 15% average issues-resolved-per-hour gain, heterogeneous experience/skill effects, customer interaction changes and discretion to ignore suggestions.
limitations
One firm's adapted technical-support assistant; field rollout differs from randomized general-office adoption. Evidence consistent with learning does not certify all unaided competence. Published 15% replaces older working-paper averages.
original id
support-qje-2025
ChatGPT and large language models: what's the risk? — UK National Cyber Security Centre (2023) ↗
id
ai-at-work--ncsc-llm-2023 — ChatGPT and large language models: what's the risk? — UK National Cyber Security Centre (2023)
kind
official
accessed at
2026-10-01
supports
Opened technical explanation of LLMs, prompts, operator access and confidential-input risk. Supports distinction between entering information and a model instantly learning it.
limitations
Dated technical guidance, not verification of current vendor retention policies. Use current approved terms for the particular service. No blanket assertion that every deployment has identical data handling.
original id
ncsc-llm-2023
AI RMF Playbook — MEASURE, NIST ↗
id
ai-at-work--nist-measure — AI RMF Playbook — MEASURE, NIST
kind
guideline
accessed at
2026-10-01
supports
Opened relevant deployment-context, validity/reliability, generalization-limit, interpretation and ongoing-feedback passages, including MEASURE 2.3, 2.5 and 2.9.
limitations
Framework guidance, not a controlled demonstration of the exact STUDlearn summary or source-comparison exercises.
original id
nist-measure
Thinking carefully before adopting agentic AI — UK National Cyber Security Centre (May 15, 2026) ↗
id
ai-at-work--ncsc-agents-2026 — Thinking carefully before adopting agentic AI — UK National Cyber Security Centre (May 15, 2026)
kind
official
accessed at
2026-10-01
supports
Opened agentic-AI explanation and recommendations on bounded actions/access, ownership, monitoring, incident response, stopping and starting with lower-stakes uses.
limitations
Organizational cyber guidance, not a current vendor setup guide or an assurance that human approval prevents every harm.
original id
ncsc-agents-2026
Experimental evidence on the productivity effects of generative artificial intelligence — Noy and Zhang, Science (2023), author-hosted manuscript ↗
id
ai-at-work--writing-2023 — Experimental evidence on the productivity effects of generative artificial intelligence — Noy and Zhang, Science (2023), author-hosted manuscript
kind
study
accessed at
2026-10-01
supports
Opened author-linked manuscript: abstract, experimental design, results and limitations. Verified 453 professionals, randomized access, approximately 40% less time and 18% higher graded quality, lower-baseline gains, attitude changes and follow-up adoption.
limitations
Paid short writing tasks using 2023 ChatGPT; no contextual knowledge or precise factual accuracy requirement. Grader scores are not client outcomes. Science publisher text was unavailable; this author's manuscript corresponds to the 453-person version. An older MIT preprint with 444 participants was excluded from numerical synthesis.
original id
writing-2023
The productivity effects of generative artificial intelligence — Noy and Zhang, VoxEU (2023) ↗
id
ai-at-work--writing-author-account — The productivity effects of generative artificial intelligence — Noy and Zhang, VoxEU (2023)
kind
professional
accessed at
2026-10-01
supports
Opened the authors' public account of writing-task design, lightly edited output, enjoyment and later reported workplace use.
limitations
An explanation of the same experiment, not independent replication or advice to send unchecked work. STUDlearn's placeholder exercise is an editorial application.
original id
writing-author-account
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST AI 600-1 (2024) ↗
id
ai-at-work--nist-gai — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST AI 600-1 (2024)
kind
guideline
accessed at
2026-10-01
supports
Opened profile and relevant risk definitions: confabulation, data privacy, information security, intellectual property and human-AI configuration including automation bias and over-reliance. Opened official publication record to verify Reva Schwartz as coauthor.
limitations
Cross-sector risk guidance rather than a prediction that every system always produces these harms. Suggested desk checks are editorial adaptations. Authorship alone does not verify a coauthor's current employment.
original id
nist-gai
The hidden risks of shadow AI — UK National Cyber Security Centre (September 7, 2026) ↗
id
ai-at-work--ncsc-shadow-2026 — The hidden risks of shadow AI — UK National Cyber Security Centre (September 7, 2026)
kind
official
accessed at
2026-10-01
supports
Opened official explanation of unapproved workplace AI use, diminished organizational visibility and information control, and the importance of usable approved routes and policy.
limitations
Cybersecurity guidance rather than a jurisdiction-specific employment rule. Unopened third-party prevalence statistics in the post were not used.
original id
ncsc-shadow-2026
How should we assess security and data minimisation in AI? — UK Information Commissioner's Office ↗
id
ai-at-work--ico-minimisation — How should we assess security and data minimisation in AI? — UK Information Commissioner's Office
kind
guideline
accessed at
2026-10-01
supports
Opened relevant security and minimisation passages: necessary personal data for a purpose, minimising processing throughout development and deployment.
limitations
UK data-protection context; site states guidance is under review following legal changes. General privacy literacy only; replacing a name alone is not a compliance determination or assured anonymisation.
original id
ico-minimisation
Formal Opinion 512: Generative Artificial Intelligence Tools — American Bar Association (July 29, 2024) ↗
id
ai-at-work--aba-512 — Formal Opinion 512: Generative Artificial Intelligence Tools — American Bar Association (July 29, 2024)
kind
guideline
accessed at
2026-10-01
supports
Opened opinion, relevant competence, confidentiality, informed consent, communication, supervision and fees passages. Reviewed limits on self-learning-tool inputs and insufficient boilerplate consent, and context-dependent client disclosure.
limitations
US lawyers and ABA Model Rules; applicability depends on jurisdiction and circumstances. Not legal advice, a general employee consent rule, or a claim that every prompt requires public disclosure.
original id
aba-512
Generative AI and Jobs: A Refined Global Index of Occupational Exposure — ILO Working Paper 140 (2025) ↗
id
ai-at-work--ilo-exposure-2025 — Generative AI and Jobs: A Refined Global Index of Occupational Exposure — ILO Working Paper 140 (2025)
kind
study
accessed at
2026-10-01
supports
Opened official UN-hosted report, author biographies, abstract and relevant methods/results. Verified task-based worker/expert/model assessment, one in four workers with some occupational exposure, 3.3% highest exposure and transformation-versus-replacement interpretation.
limitations
Modeled potential exposure rather than observed job loss or employer-specific forecast. Employment distributions and the task-rating procedure shape estimates. Relevant portions reviewed, not every page. ILO publication landing page was also opened.
original id
ilo-exposure-2025
Why Are There Still So Many Jobs? The History and Future of Workplace Automation — David Autor, Journal of Economic Perspectives 29(3), 3–30 (2015) ↗
id
ai-at-work--autor-2015 — Why Are There Still So Many Jobs? The History and Future of Workplace Automation — David Autor, Journal of Economic Perspectives 29(3), 3–30 (2015)
kind
review
accessed at
2026-10-01
supports
Opened official abstract explaining substitution, complementarity and changing demand as reasons task automation does not directly determine total employment.
limitations
Historical conceptual synthesis, not a 2026 generative-AI employment forecast. Full paper was not reviewed in this pass; application to a three-line task card is editorial.
original id
autor-2015
Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua and colleagues, Organization Science (2026) ↗
id
ai-at-work--frontier-2026 — Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua and colleagues, Organization Science (2026)
kind
study
accessed at
2026-10-01
supports
Opened final journal article on publisher and HBS-hosted PDF: design and results for 758 BCG consultants; 18 in-frontier tasks; 12.2% more tasks, 25.1% faster, higher-rated quality and larger lower-baseline gains. Checked outside-frontier group correctness supporting approximately 19 percentage points lower correctness across AI conditions.
limitations
GPT-4 in a specified consultant experiment, not a map of current capability. Final article published online March 11, 2026, DOI 10.1287/orsc.2025.21838. Do not combine its estimates with the 2023 working paper's 40% quality headline or treat percentage-point correctness difference as a universal relative-percent effect.
original id
frontier-2026
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METR ↗
id
ai-at-work--metr-2025 — Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METR
kind
study
accessed at
2026-10-01
supports
Opened research organization's study report: randomized AI access for 16 experienced developers and 246 real repository tasks; 19% slower completion and perceived speedup.
limitations
Small, selected sample, familiar complex repositories and early-2025 tools. Organization's report was reviewed rather than an independent replication. Not all coding or current models.
original id
metr-2025
Update on AI developer productivity — METR (February 24, 2026) ↗
id
ai-at-work--metr-update-2026 — Update on AI developer productivity — METR (February 24, 2026)
kind
study
accessed at
2026-10-01
supports
Opened follow-up report and its discussion of participation/task selection, changed compensation and time measurement with concurrent agents. Authors consider greater uplift likely but say later estimates are unreliable for determining its size.
limitations
Not a clean measured reversal of the 2025 result. Selection and measurement changes prevent a strong contemporary productivity estimate.
original id
metr-update-2026
One in four jobs at risk of being transformed by GenAI, new ILO–NASK Global Index shows — ILO ↗
id
ai-at-work--ilo-brief-2025 — One in four jobs at risk of being transformed by GenAI, new ILO–NASK Global Index shows — ILO
kind
official
accessed at
2026-10-01
supports
Opened official explanation of the refined 2025 index, variation by occupational composition, income group and sex, and potential transformation.
limitations
Public explanation of the same index, not an additional employment study. Risk of transformation is not a count of layoffs.
original id
ilo-brief-2025
AI Companies: Uphold Your Privacy and Confidentiality Commitments — FTC Office of Technology (January 2024) ↗
id
ai-at-work--ftc-privacy-2024 — AI Companies: Uphold Your Privacy and Confidentiality Commitments — FTC Office of Technology (January 2024)
kind
official
accessed at
2026-10-01
supports
Opened staff post explaining potential liability for providers breaking privacy/confidentiality commitments, including promises not to train on customer data.
limitations
Provider-facing staff guidance, not permission for an employee to upload a file. Current individual provider terms were not audited. No-training and no-transfer are different issues.
original id
ftc-privacy-2024
How generative AI 'persuasion bombs' users — and how to fight back — MIT Sloan (April 28, 2026) ↗
id
ai-at-work--mit-validation-2026 — How generative AI 'persuasion bombs' users — and how to fight back — MIT Sloan (April 28, 2026)
kind
professional
accessed at
2026-10-01
supports
Opened institutional research explanation of qualitative interaction analysis involving 72 BCG consultants, GPT-4 and 4,339 prompts; discusses increasingly persuasive responses to challenge and verification outside the chat.
limitations
The institutional account was reviewed; the full working paper was unavailable. It describes one experimental setting, not intentional deceit, every model or a universal rate.
original id
mit-validation-2026
AI RMF Playbook — GOVERN, NIST ↗
id
ai-at-work--nist-govern — AI RMF Playbook — GOVERN, NIST
kind
guideline
accessed at
2026-10-01
supports
Opened relevant policies, responsibility, accountability, training, feedback and oversight suggestions; reviewers need ability to challenge or override results.
limitations
Suggested actions adapted to context, not all mandatory steps or workplace legal requirements. No claim that a policy alone establishes correct output.
original id
nist-govern
Artificial Intelligence Risk Management Framework (AI RMF 1.0) — NIST AI 100-1 (2023) ↗
id
ai-at-work--nist-rmf — Artificial Intelligence Risk Management Framework (AI RMF 1.0) — NIST AI 100-1 (2023)
kind
guideline
accessed at
2026-10-01
supports
Opened framework and relevant scope, trustworthiness, risk, Core and human-role passages. Supports voluntary nature, likelihood/severity, seven trustworthiness characteristics and four connected functions.
limitations
Voluntary framework, not binding law, a certification, or a validated personal checklist. Framework concepts do not establish an organization's compliance.
original id
nist-rmf
AI RMF Playbook — MANAGE, NIST ↗
id
ai-at-work--nist-manage — AI RMF Playbook — MANAGE, NIST
kind
guideline
accessed at
2026-10-01
supports
Opened relevant risk-response, monitoring, error/incident communication, appeal and override passages including MANAGE 4.1 and 4.3.
limitations
Context-specific suggested actions; does not create universal worker rights or prescribe a two-minute personal review.
original id
nist-manage
How to use generative AI to augment your workforce — MIT Sloan (February 24, 2025) ↗
id
ai-at-work--mit-define-success-2025 — How to use generative AI to augment your workforce — MIT Sloan (February 24, 2025)
kind
professional
accessed at
2026-10-01
supports
Opened Danielle Li's public implementation guidance on defining success, data/infrastructure, workforce effects and evaluating augmentation.
limitations
Professional explanation; not randomized validation of STUDlearn's small workflow comparison or guaranteed return on a tool.
original id
mit-define-success-2025
10 levers shaping whether generative AI truly improves worker performance — MIT Sloan (September 3, 2026) ↗
id
ai-at-work--mit-levers-2026 — 10 levers shaping whether generative AI truly improves worker performance — MIT Sloan (September 3, 2026)
kind
professional
accessed at
2026-10-01
supports
Opened recent institutional synthesis stressing context, worker judgment, evidence before scaling, knowing when to trust output and costs of rework.
limitations
Professional synthesis, not a single causal estimate for every lever or permission for autonomous work.
original id
mit-levers-2026
AI RMF Core — NIST AI Resource Center ↗
id
ai-at-work--nist-core — AI RMF Core — NIST AI Resource Center
kind
guideline
accessed at
2026-10-01
supports
Opened Core: GOVERN, MAP, MEASURE, MANAGE, human oversight, responsibilities, training, affected-party feedback and the statement that functions are not a required ordered checklist.
limitations
Organizational voluntary guidance. Naming a human does not establish meaningful review or a uniform legal right of appeal.
original id
nist-core
The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers — Lane, Williams and Broecke (2023) ↗
id
ai-at-work--oecd-surveys-2023 — The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers — Lane, Williams and Broecke (2023)
kind
study
accessed at
2026-10-01
supports
Opened official publication overview and report; verified manufacturing/finance, seven countries, positive reported outcomes with job-loss concerns, and training/consultation associations. Sample/timing additionally verified on current OECD survey project page.
limitations
Early-2022 perception surveys, not a causal evaluation, a study of current ChatGPT adoption or a survey of all sectors. Relevant report sections were reviewed, not every questionnaire and table. Association does not establish a universal training effect.
original id
oecd-surveys-2023
OECD AI surveys of employers and workers — project overview ↗
id
ai-at-work--oecd-survey-design — OECD AI surveys of employers and workers — project overview
kind
official
accessed at
2026-10-01
supports
Opened About text verifying early-2022 sample of 5,334 workers and 2,053 firms in Austria, Canada, France, Germany, Ireland, UK and US. States surveys repeated in 2026 with wider coverage.
limitations
2026 repetition is not treated as published outcome evidence; this article reports the 2023 publication of 2022 data. Project overview is not an independent study.
original id
oecd-survey-design
Kate Kellogg — MIT Sloan faculty directory ↗
id
ai-at-work--professional-kellogg — Kate Kellogg — MIT Sloan faculty directory
kind
professional
accessed at
2026-10-01
supports
Opened faculty biography verifying management/innovation role and public organizational research on technology, work practices and implementation; consulted related official MIT explainers listed above.
limitations
Credential and research-context verification, not empirical proof that a specific training or logging scheme will work.
original id
professional-kellogg
Ifeoma Ajunwa — Emory Law faculty profile ↗
id
ai-at-work--professional-ajunwa — Ifeoma Ajunwa — Emory Law faculty profile
kind
professional
accessed at
2026-10-01
supports
Opened biography identifying Asa Griggs Candler Professor of Law, AI/Future of Work initiative leadership and scholarship on quantified workers, employment technology, privacy and discrimination.
limitations
Professional context distinguishing worker assessment from assistance with one's own draft. Profile is not jurisdiction-specific legal advice or a reviewed causal effect of monitoring.
original id
professional-ajunwa
Humans in the Loop: The Evolution of Work in Early Experiments with Generative AI — Armstrong and Shah, MIT Industrial Performance Center (April 2026) ↗
id
ai-at-work--mit-work-design-2026 — Humans in the Loop: The Evolution of Work in Early Experiments with Generative AI — Armstrong and Shah, MIT Industrial Performance Center (April 2026)
kind
study
accessed at
2026-10-01
supports
Opened report executive summary, methods and relevant findings on early experiments at more than 20 companies in four industry groups, variation in supervisory work and job quality.
limitations
Qualitative interviews/cases, not causal net-employment estimates or representative universal outcomes. Relevant portions reviewed rather than every detail.
original id
mit-work-design-2026
Generative AI without guardrails can harm learning: Evidence from high school mathematics — Bastani and colleagues, PNAS 122(26), e2422633122 (2025) ↗
id
ai-at-work--bastani-pnas-2025 — Generative AI without guardrails can harm learning: Evidence from high school mathematics — Bastani and colleagues, PNAS 122(26), e2422633122 (2025)
kind
study
accessed at
2026-10-01
supports
Opened published article methods/results: field experiment in Turkish high-school math, GPT-4-based unrestricted and safeguarded tutors, assisted practice improvements, lower later unaided performance with unrestricted helper and mitigation with safeguarded design.
limitations
Classroom mathematics, not adult workplace evidence. Mitigating the unaided penalty is not demonstrating a later learning gain. Use published version rather than imply that the earlier working paper remains the latest evidence.
original id
bastani-pnas-2025
The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — Lee, Sarkar, Tankelevitch and colleagues, CHI (2025) ↗
id
ai-at-work--critical-thinking-2025 — The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — Lee, Sarkar, Tankelevitch and colleagues, CHI (2025)
kind
study
accessed at
2026-10-01
supports
Opened author institution's publication record and abstract: 319 workers, 936 examples, confidence associations and shift toward verification, integration and task stewardship; DOI 10.1145/3706598.3713778.
limitations
Self-reports and associations, not an objective causal test of skill loss. Detailed full-paper analyses were not reviewed. Microsoft-employed authors' institutional affiliation is disclosed in panel.
original id
critical-thinking-2025
What do we need to know about accuracy and statistical accuracy? — UK Information Commissioner's Office ↗
id
ai-at-work--ico-accuracy — What do we need to know about accuracy and statistical accuracy? — UK Information Commissioner's Office
kind
guideline
accessed at
2026-10-01
supports
Opened distinction between accuracy of personal data and statistical accuracy of models; relevant guidance on marking predictions/inferences rather than presenting them as facts.
limitations
UK legal/regulatory context, under review. Does not determine whether one workplace record is lawful or prescribe a universal record format.
original id
ico-accuracy
ExperiencedDevs discussion of the early-2025 developer productivity study ↗
id
ai-at-work--question-speed-perception — ExperiencedDevs discussion of the early-2025 developer productivity study
kind
question_signal
accessed at
2026-10-01
supports
Opened discussion to identify a reader question: why can assistance feel faster while measured completion is slower? The answer uses METR's opened reports.
limitations
Anonymous anecdotes and thread popularity are not evidence of productivity, professional consensus or a validated workflow. No factual guidance relies solely on this discussion.
original id
question-speed-perception
Futurology discussion of a headline about AI and thinking skills ↗
id
ai-at-work--question-thinking-headline — Futurology discussion of a headline about AI and thinking skills
kind
question_signal
accessed at
2026-10-01
supports
Opened discussion as a signal of confusion between a confidence/self-report association and a claim that AI destroys thinking. Article explains the actual CHI survey's bounds.
limitations
Headline and comments are not causal evidence. No X access or professional consensus is claimed.
original id
question-thinking-headline

Professional perspectives

Erik Brynjolfsson
Stanford Jerry Yang and Akiko Yamazaki Professor; Stanford Digital Economy Lab director, as listed on Stanford Economics profile
Public usefulness
Public digital-economy research, books and explanations help readers distinguish an aggregate productivity claim from a specific workplace deployment.
Contribution
Coauthored the published support-agent study used for average and heterogeneous effects.
Scope and limits
That study covers one firm's support assistant; his broader expertise does not establish every tool's effects.
Reviewed
2026-10-01
Danielle Li
David Sarnoff Professor of Management of Technology, MIT Sloan, as listed by MIT
Public usefulness
MIT's public augmentation guidance explains how to define success before adopting workplace AI.
Contribution
Coauthored support-agent evidence and supplies the practical success-criteria framing.
Scope and limits
Institutional implementation advice does not validate STUDlearn's exact example or promise savings.
Reviewed
2026-10-01
Lindsey Raymond
Assistant Professor jointly in MIT Economics, EECS and the Schwarzman College of Computing, as listed by MIT
Public usefulness
Public research on AI, labor and human learning makes workplace effects understandable through actual data rather than demonstration alone.
Contribution
Coauthored Generative AI at Work, including experience differences and evidence consistent with learning.
Scope and limits
Profile is no longer described as an incoming appointment; the research still does not establish universal learning or employment effects.
Reviewed
2026-10-01
Fabrizio Dell'Acqua
HBS AI Institute researcher with a Columbia Management PhD; official personal site lists an incoming Michigan Ross assistant professorship
Public usefulness
Public research explanations and teaching describe why nearby-looking knowledge tasks can produce different AI effects.
Contribution
Lead author of the consultant frontier experiment, checked in its 2026 journal version.
Scope and limits
Incoming appointment is reported as listed, not asserted as already begun; task boundaries from the experiment are not today's universal capability map.
Reviewed
2026-10-01
Ethan Mollick
Associate Professor of Management and co-director of the Generative AI Labs at Wharton, as listed by Knowledge at Wharton
Public usefulness
Public explanations and Co-Intelligence help ordinary readers consider AI assistance and human judgment together.
Contribution
Coauthored the consultant experiment supporting uneven task performance.
Scope and limits
Public examples and personal tool experiments do not establish a prompt's reliability or replace independent checking.
Reviewed
2026-10-01
Hamsa Bastani
Associate Professor of Operations, Information and Decisions at Wharton with a secondary Statistics and Data Science appointment, as listed by Wharton
Public usefulness
Public research on human/AI decision systems and safeguarded tutoring clarifies assisted performance versus later learning.
Contribution
Coauthored the published high-school mathematics trial comparing unrestricted and safeguarded GPT-4 tutors.
Scope and limits
Classroom learning evidence is adjacent to adult workplace accountability; it does not prove workplace deskilling.
Reviewed
2026-10-01
David Autor
Professor of Economics at MIT, verified through MIT faculty profile
Public usefulness
Accessible essays explain tasks, substitution, complementarity and demand as mechanisms of workplace change.
Contribution
2015 automation review supplies the conceptual distinction between automating a task and eliminating an entire occupation.
Scope and limits
Historical framework, not a current numerical employment forecast or reassurance against displacement.
Reviewed
2026-10-01
Paweł Gmyrek
Senior Researcher in the ILO Research Department's Effective Labour Institutions work, as listed in ILO research repository
Public usefulness
Public ILO reports explain occupational exposure in a way readers can separate from layoffs.
Contribution
Lead author of the refined 2025 ILO–NASK exposure index.
Scope and limits
Task ratings and exposure estimates concern potential change, not observed employer-specific outcomes.
Reviewed
2026-10-01
Janine Berg
ILO Research Department Senior Economist, identified in the opened May 2025 report's author biography
Public usefulness
ILO's public exposure reports connect technological change to working conditions and transition support.
Contribution
Coauthored the refined exposure index and its transformation-versus-replacement interpretation.
Scope and limits
Role is verified in the dated report biography, not independently confirmed as a current appointment; modeling does not establish net job loss.
Reviewed
2026-10-01
Elham Tabassi
Director of the Artificial Intelligence and Emerging Technology Initiative and Senior Fellow at Brookings; former NIST chief AI advisor and AI RMF leader, as listed by Brookings
Public usefulness
Public risk-framework work explains responsibility, trustworthy systems and meaningful measurement without relying on vendor marketing.
Contribution
Led the AI RMF that supplies the four-function and trustworthiness teaching.
Scope and limits
Current Brookings role is distinguished from historical NIST leadership; voluntary framework is not law or certification.
Reviewed
2026-10-01
Arvind Narayanan
Princeton Computer Science professor and Center for Information Technology Policy director, as listed on official faculty page
Public usefulness
Public AI as Normal Technology explanations encourage readers to distinguish demonstrated capabilities from advertised possibilities.
Contribution
Provides the public explanatory lens of evaluating what a system actually does; factual risk claims here rest on opened studies and NIST guidance.
Scope and limits
Profile and public teaching are professional context, not validation of one employer's assistant or evidence of a universal model failure rate.
Reviewed
2026-10-01
Ifeoma Ajunwa
Asa Griggs Candler Professor of Law at Emory and founding director of its AI and the Future of Work initiative, as listed by Emory
Public usefulness
The Quantified Worker and public employment-technology scholarship help readers ask about privacy, monitoring and decisions affecting workers.
Contribution
Keeps AI used to evaluate a worker distinct from AI chosen by a worker to draft a message.
Scope and limits
Professional scholarship and profile are not jurisdiction-specific legal advice or a universal right-of-appeal claim.
Reviewed
2026-10-01
Stijn Broecke
Senior Economist in OECD Employment, Labour and Social Affairs, leading future-of-work research, as listed by OECD.AI
Public usefulness
Public OECD research explanations and talks connect workforce consultation and training to reported experiences.
Contribution
Coauthored the 2023 AI workplace survey report used for training/consultation associations.
Scope and limits
Early-2022 survey perceptions are not a randomized causal effect or a survey of current generative tools.
Reviewed
2026-10-01
Kate Kellogg
David J. McGrath Jr. Professor of Management and Innovation at MIT Sloan, as listed by MIT
Public usefulness
Public organizational research and MIT explainers describe how implementation, expertise and worker participation change technology's practical effects.
Contribution
Coauthored consultant frontier research and contributes organizational context for training, work design and review.
Scope and limits
Neither a profile nor implementation research guarantees a particular training package's causal effect.
Reviewed
2026-10-01
Reva Schwartz
Named coauthor of NIST AI 600-1, the Generative Artificial Intelligence Profile (2024), verified on NIST's publication record
Public usefulness
Coauthored publicly accessible professional risk guidance explaining confabulation, privacy and human-AI configuration.
Contribution
Contributes to the official profile supporting over-reliance and confidentiality risk teaching.
Scope and limits
Professional authorship is verified; a current job was not independently confirmed. Coauthorship does not attribute every passage to this individual.
Reviewed
2026-10-01
Advait Sarkar
Senior Researcher, Microsoft Research Cambridge, as listed on Microsoft Research profile
Public usefulness
Public human-computer-interaction research helps readers understand verification and critical-thinking effort in AI-assisted work.
Contribution
Coauthored the CHI survey of 319 workers and 936 examples used to clarify confidence associations.
Scope and limits
Microsoft-employed researcher; self-reported associations are not causal skill-loss evidence. Profile warns it is not regularly updated, so role is stated as listed.
Reviewed
2026-10-01
Lev Tankelevitch
Senior Researcher and behavioral scientist at Microsoft Research Cambridge; Oxford PhD background in cognitive psychology/neuroscience, verified on Microsoft profile
Public usefulness
Public metacognition and human-agency explanations connect thinking about one's own judgment to AI-assisted collaboration.
Contribution
Coauthored the critical-thinking survey and supplies relevant expertise for confidence versus verification distinctions.
Scope and limits
Microsoft-employed researcher. Survey supports reported associations, not a universal deskilling effect or a validated workplace intervention.
Reviewed
2026-10-01
Shakked Noy
MIT Economics PhD student, described on official personal site as fifth-year and on the 2026–27 job market
Public usefulness
Public coauthored VoxEU account explains writing-task design and the gap between an experimental result and workplace use.
Contribution
Coauthored the 453-person Science writing experiment; the author's site links the manuscript used here.
Scope and limits
The listed PhD is in progress. The short, generic writing tasks do not establish real-file accuracy or productivity across workplaces.
Reviewed
2026-10-01

Keep this topic in view

Preserved original record

The earlier public profile is preserved separately from the current article. It may contain superseded wording and is not newly checked material.

Read the full original profile

slug

ai-at-work

name

AI at work

browse Line

Use workplace AI with human judgment, careful checking, and privacy in mind.

photo

/learn/assets/topics/ai-at-work.jpg

photo Alt

An adult at a home desk looking at a laptop before continuing to type

aliases

ChatGPT work
Copilot at work
workplace AI
generative AI job
paste into chatbot
AI draft
AI replace my job
confidential ChatGPT
AI email
job automation

family

work

heat

mid

talk

Sources

most Discussed

true

first Line

Workplace generative AI produces drafts, summaries and other outputs from instructions and supplied material. This topic covers where it can help, how to check the work and how to handle information responsibly.

why Line

feel Words

draft
email
report
write
ChatGPT
Copilot
summarize
check

common Picture

lead
Start a work draft without handing over the job
why
A chatbot can produce a fluent first pass of an email, summary, or short report while you still own the facts, the tone, and the send button. Generative AI means a system that creates new text, images, or code from a prompt rather than retrieving a stored document. Shakked Noy and Whitney Zhang ran a preregistered online experiment with 453 college-educated professionals on occupation-specific writing tasks. People given ChatGPT finished faster and received higher grader scores on those tasks; many pasted the prompt and submitted lightly edited output. That is a bounded writing experiment, not a guarantee for your workplace, your clients, or every kind of document. Try asking for a structure or a public-style rewrite, then put the real names, numbers, and judgment back in yourself.
lead
Check a fluent draft before anyone relies on it
why
A polished paragraph can still invent a date, a citation, or a policy that was never there. NIST’s generative-AI profile calls this confabulation: confidently stated but erroneous content. Fabrizio Dell’Acqua and colleagues found that Boston Consulting Group consultants using GPT-4 did better on tasks inside the tool’s current reach and were 19 percentage points less likely to produce a correct solution on a task chosen to sit outside it. The authors call that uneven pattern a jagged technological frontier: nearby-looking jobs are not equally doable. Before you send, compare the draft with a source you own — the original email, the current notice, or the spreadsheet — and keep the check proportionate to the cost of being wrong.
lead
Keep confidential work out of tools you do not control
why
Pasting a client name, a contract, a patient detail, or an unreleased figure into a public chatbot can be a disclosure even if you later delete the chat. Confidentiality means a duty not to share information relating to a person, a client, or an employer except as authorized. The American Bar Association’s Formal Opinion 512 tells lawyers to understand how a generative tool uses inputs and to obtain informed consent before putting information relating to a representation into a self-learning tool. NIST treats privacy-enhanced design as a trustworthiness characteristic, and the FTC has reminded AI providers to keep confidentiality commitments. Professional and contractual duties differ by role. If your workplace, license, or contract restricts sharing, keep identifiers and secret material out unless an approved tool and a lawful basis already cover that use.
lead
Divide the job into tasks you still own
why
A job combines tasks that differ in how much information, judgment and responsibility they require. The ILO–NASK 2025 index estimates potential exposure to generative AI by examining tasks within occupations; it does not count jobs already lost. Its conclusion still favors transformation of many roles over complete replacement. A useful workplace discussion therefore names the task: drafting an outline, checking a calculation, answering a client or making a consequential decision. Ask what the tool can do reliably, who checks the result and which responsibility remains with a person. Adoption and job redesign determine what potential exposure becomes in practice.

situation Quick Answers

- label
  A useful first task
  value
  Ask for a structure or a rewrite using public material.
- label
  Your part
  value
  Put the correct names, numbers, tone, and judgment back into the draft before sending.
- label
  Before anyone relies on it
  value
  Compare the draft’s facts and citations with the original email, current notice, spreadsheet, or other source you own.
- label
  How much checking?
  value
  Match the review to the cost of being wrong; fluent writing can still be incorrect.
- label
  Use an approved route
  value
  Put restricted work into a tool only when workplace rules, professional duties, and authorization already cover that use.
- label
  Otherwise
  value
  Keep client identifiers, patient details, contracts, and secret figures out.
- label
  Divide one job
  value
  Name one task AI may help with, one check you still perform, and one part that requires your judgment, access, or sign-off.
- label
  What changes
  value
  Some tasks can change without the whole occupation disappearing.

picture Lead

Start a work draft without handing over the job

watch For

what People Say

Separate a task’s exposure from a job’s future. The ILO–NASK 2025 index estimates where generative AI could affect occupational tasks. A high exposure score is not a prediction that every worker in that occupation will lose a job. Jobs combine automatable steps with coordination, accountability and work that a tool cannot reliably complete. Use the estimate to ask which tasks may change and what support workers need, rather than treating an occupation label as a forecast.
Use a draft as a starting point, then apply your own check. Noy and Zhang assigned realistic writing tasks to managers, marketers, grant writers, consultants, human-resource professionals and data analysts. Access to ChatGPT reduced time and raised grader scores on those mid-level writing tasks; inequality between participants narrowed because lower-scoring writers gained more. Their own VoxEU account notes that many treated participants submitted unedited or lightly edited chatbot text, and that graders liked that text. The experiment cannot promise the same result in a high-stakes file, a regulated filing, or a document whose facts live only in your workplace. Keep the useful part: a first pass can save typing, while the sender still supplies facts, tone, and a check.
Expect uneven help across workers and tasks. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied a generative conversational assistant rolled out to 5,172 customer-support agents. Access raised issues resolved per hour by about 15 percent on average, with larger gains for less experienced and lower-skilled agents and small speed gains plus small quality declines among the most experienced. Agents could accept or reject suggestions; they were not required to follow every line. That is one firm’s support chats, not every office. Kate Kellogg’s public explanation of related knowledge-work experiments is that a tool can level some skills on tasks it can already do. Plan for difference by task and by experience rather than a uniform boost.
Protect confidential material before you type. A prompt is a disclosure to the operator of the tool unless a contract, approved system, and your workplace rules already cover that transfer. ABA Formal Opinion 512 treats confidentiality as applying to information relating to a representation regardless of its source, and it warns that self-learning tools can raise disclosure risk even inside one firm. The ICO’s AI and data-protection guidance keeps data minimisation and a lawful basis in view when personal data are processed. NIST’s privacy-enhanced characteristic asks whether people retain control over identity and confidential facts. A practical workplace adaptation is to draft with placeholders, public facts, or an approved tool — not a consumer chatbot — when the material is someone else’s.
Treat fluent output as something to verify. Confabulation is NIST’s term for confidently stated but erroneous generative content, including answers that drift from the prompt or contradict an earlier line. Validity and reliability in the AI Risk Management Framework mean a system does what it is intended to do under expected conditions; fluency is not that test. Arvind Narayanan’s public teaching on AI that does not work as advertised is a reason to ask what the draft is actually claiming. Compare names, dates, figures and citations with a record you control. The amount of checking should fit the consequence: a private outline needs less than a client email, a public notice, or a licensed filing.
Map the risk of this use, then decide how to check it. NIST’s AI Risk Management Framework is voluntary guidance for organizations and individuals who design, deploy or use AI. Its four functions are GOVERN, MAP, MEASURE and MANAGE: set roles and rules, describe the context and possible harms, test what you can, and respond. Trustworthiness here includes being valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. Elham Tabassi led that framework. You do not need a full enterprise program to use the idea on one task: name who is accountable for the send, what could go wrong, and what check you will actually perform.

disputed Notes

Match training and consultation to how the tool is introduced. OECD surveys of 5,334 workers and 2,053 firms in manufacturing and finance across seven countries found that many AI users said performance and some working-condition measures had improved, while job-loss worry remained. Stijn Broecke, Marguerita Lane and Morgan Williams report that training and worker consultation were associated with better reported outcomes; that is a survey association, not a proof that one training package causes those results. Employers and workers can reasonably disagree about how fast to adopt a tool, who is consulted, and how intensity or monitoring might change. Ask what training exists, who was asked, and what is still uncertain rather than treating either enthusiasm or worry as the whole field.
Match AI use to the task and the worker’s starting point. The consultant and customer-support studies do not show opposite novice effects. Dell’Acqua and colleagues found larger gains for lower-baseline performers on tasks within GPT-4’s capability, while Brynjolfsson and colleagues found larger gains for novice and less-skilled support agents. They studied different work, tools and settings. The consultant experiment also found worse correctness on a deliberately difficult task outside the tool’s capability. Expertise, task design and verification therefore remain relevant; neither result says an inexperienced worker can safely delegate every judgment.
Choose whether and how to name AI use. Some workplaces require disclosure of generative-AI assistance; some clients or professions treat it as part of competent work; some treat undisclosed use as a problem of candor or authorship. ABA Formal Opinion 512 discusses when lawyers should consult clients about the means used, including generative tools, without turning every prompt into a public announcement. Ifeoma Ajunwa’s workplace-technology work keeps a different disclosure problem in view: automated hiring and monitoring tools can affect workers who never chose them. There is no single reviewed rule for every email. Follow the policy that governs your role, and when the output will be relied on by someone else, make the human review visible even if the tool is not named.
Treat employment forecasts as scenarios with assumptions. Task exposure can inform forecasts, but it does not settle employer adoption, customer demand or how saved time will be used. The ILO’s refined 2025 index explicitly concerns potential transformation. Firms may reorganize jobs differently even when they buy similar tools. When comparing a prediction with workplace experience, check its date, region, occupation and whether it measured actual employment or modeled technical capability.

talk Notes

Judge a tool by later work you can still do. Speed during a session is not the same as competence once the helper is closed. Hamsa Bastani and colleagues found that a ChatGPT-style GPT-4 tutor helped high-school math students while it was available, then left them worse than a no-AI group on later unaided tests; a more constrained tutor largely avoided that drop. That is a classroom finding, not a workplace ban. At work the adjacent question is whether you can still explain, check, or redo the part of the job you remain responsible for. Keep a later look without the chat open when the role still requires that judgment.
Match the check to what is at stake. Two extra minutes on a routine outline is not the same as sending an unchecked number to a client. NIST asks organizations to consider both likelihood and magnitude of harm, including harm to people, organizations and broader systems. Narayanan’s distinction between systems that work as advertised and systems that do not is useful here: a convincing tone is not evidence. Write what would actually go wrong if a sentence were false, then choose a check that can catch that error. Popular talk that “the model already checked it” skips that step.
Keep the duty of confidentiality even after you delete a chat. Deleting a conversation does not unsend a disclosure. ABA Formal Opinion 512 treats the input itself as the confidentiality event. The FTC has told AI providers to honor promises about how customer data are used, including commitments not to train on that data; those promises bind the company, they do not rewrite your contract with a client or employer. ICO guidance still asks for a lawful basis and data minimisation when personal data are processed. If you would not email the material to a stranger, keep it out of a consumer tool.
Keep the skill the job still asks of you. A completed chatbot draft is not proof that you can do the work unaided, supervise it, or sign it. Autor’s task framework and the ILO exposure analysis both separate automatable steps from remaining occupational work. Ajunwa’s research on quantified work and automated hiring is a reminder that a score or a generated paragraph can hide the human decision. Practice the part you still own: the facts, the exception, the conversation, or the approval. A slogan that “AI means I no longer need the knowledge” is popular talk, not a finding.

how People Run It

Name one task, keep private material out, then check the draft. Write the smallest work task you actually want help with this week, such as structuring a public-style email or listing questions for a meeting. Confirm that the prompt contains no client, patient, student, employee or employer-confidential detail. Ask the approved tool, if you have one, or a consumer tool only with public or placeholder facts. Then compare the result with a source you own and correct one concrete miss. Noy and Zhang support a bounded writing-task experiment; ABA Formal Opinion 512 and NIST privacy guidance support keeping confidential inputs out. This everyday sequence is an adaptation, not their procedure and not a productivity target.
Ask for a structure using placeholders, then fill the facts yourself. Replace real names with Role A, Date, and Amount. Request an outline or a tone pass, not a finished message that pretends to know your file. Noy and Zhang observed that many experimental writers submitted nearly raw chatbot text; your job is the opposite of that shortcut when the facts are yours. Check the outline against the original request. If the structure helps, keep it. If it invents a commitment, delete that line rather than polishing it.
Compare one generated claim with a record you control. Pick a date, figure, policy name or citation in the draft and open the spreadsheet, notice or source it is supposed to reflect. NIST’s confabulation risk and Dell’Acqua’s outside-frontier result are reasons to do this even when the sentence sounds finished. Write “matches,” “wrong,” or “not in the source.” That three-way label is more useful than a vibe about quality. For a high-stakes send, have a second person or the accountable owner look at the same comparison.
Write which part of the job you still own after the draft. On a card, list the tool’s contribution, your check, and the remaining human step: a conversation, an exception, an approval, or a licensed sign-off. ILO task-exposure work and Autor’s automation essay support dividing jobs into tasks rather than treating the occupation as a single switch. Brynjolfsson, Li and Raymond’s agents still decided whether to use a suggestion. If you cannot name the remaining step, the tool is being asked to own too much of the work.

how People Keep It

Keep the workplace policy beside the tool. An approved system, a banned consumer app, and an unwritten habit are different situations. OECD findings associate consultation and training with better reported worker outcomes; they do not replace the page that actually governs your role. Save the current policy link or the named owner of the approved tool. If there is no written rule yet, treat confidential material as off-limits in public tools and ask before expanding use. Policy is a constraint, not a substitute for checking the output.
Keep confidential names and numbers out of the prompt. Identifiers, health or student records, unreleased financials, source code, credentials and privileged legal material do not belong in a consumer chatbot. ABA Formal Opinion 512, ICO data-minimisation guidance and NIST’s privacy characteristic all point at control of inputs. Use placeholders or an approved environment. When in doubt, leave the detail out and add it after the draft exists on a system you are allowed to use.
Keep a checkable source next to the draft. A generated paragraph needs the email, notice, spreadsheet or handbook it is supposed to reflect. NIST’s valid-and-reliable characteristic means a system does what it is intended to do under expected conditions; eloquence is not that test. Attach the source or a short locator so a later reader, including you, can repeat the check. Where two sources disagree, keep the disagreement visible instead of letting the fluent sentence pick a winner.
Keep a later look without the chatbot open. Close the chat and explain the decision, the number, or the next step in your own words. Bastani’s classroom result is only adjacent evidence: unaided later performance can fall when a helper did the thinking. The workplace version is accountability for work you still have to stand behind. If you cannot reconstruct the work, you are not ready to send or to supervise it. For the study-and-memory version of that distinction, see Learning and remembering.
Keep a short note of what you asked, without secrets. A one-line record of the task, the tool, and whether you checked the output is enough for many ordinary jobs. Kellogg’s implementation research emphasizes that tools work differently depending on how people are included and trained. A secret prompt log that contains client data creates a new copy of the confidential material. Store only what you are allowed to keep, and match any monitoring or logging to workplace rules rather than inventing a personal surveillance file.

claims Notes

Use a draft as a starting point, then check the facts. Claim: “ChatGPT-style tools can change time and quality on some mid-level writing tasks in a bounded experiment” — established. Noy and Zhang found shorter times and higher grader scores on occupation-specific writing tasks among 453 professionals given ChatGPT. That supports a bounded experimental claim, not a promised workplace saving, a client-ready file, or every kind of document.
Divide the job into tasks instead of reading a career fate. Claim: “Generative AI will replace every job this year” — not what the research found. The ILO–NASK 2025 index estimates task exposure, not a count of jobs that will disappear. A job combines several kinds of work, and actual employment effects depend on adoption, organizational choices and demand. Use the index to identify likely areas of change rather than to declare an entire occupation obsolete.
Keep confidential material out of tools you do not control. Claim: “Confidential client or employer material is safe in a public chatbot if I delete the chat” — not what the research found. The disclosure happens when the information is entered. ABA Formal Opinion 512 treats confidentiality as attaching to inputs; ICO guidance still requires a lawful basis and minimisation for personal data; FTC staff have reminded providers to keep confidentiality promises. Deleting a thread does not unsend the transfer.
Verify fluent output against a source you own. Claim: “If the chatbot wrote it fluently, it is ready to send” — not what the research found. NIST calls confidently wrong generative content confabulation. Dell’Acqua found consultants using GPT-4 were less likely to be correct on a task outside the tool’s then-current reach. Fluency is not validity. Check names, dates, figures and citations before anyone relies on the draft.
Expect uneven help by task and experience. Claim: “A workplace AI assistant raises every worker’s performance equally” — not what the research found. Brynjolfsson, Li and Raymond found larger gains for less experienced support agents and small quality declines among the most experienced. Dell’Acqua found larger quality jumps for lower baseline consultants on in-frontier tasks and worse correctness outside the frontier. Uneven results are the documented pattern, not a uniform boost.
Ask what training and consultation actually exist. Claim: “Training and consulting workers about AI is linked with better reported outcomes” — disputed. OECD surveys associated training and consultation with better reported worker outcomes in manufacturing and finance in seven countries. That is an association in perception data, not a randomized proof that one program works everywhere. It is still a practical question to ask.
Keep the skill and judgment the job still requires. Claim: “Using AI at work means I no longer need to know the material” — popular talk. A completed draft is not unaided competence. Bastani’s classroom result is adjacent evidence that a helper can become a crutch; ILO and Autor keep remaining human tasks in view. If you still have to explain, check or sign the work, you still need the knowledge.

deep Dive

Separate a writing-task result from a workplace guarantee. Noy and Zhang paid professionals for two occupation-specific writing tasks and randomized ChatGPT access on the second. Time fell by about 40 percent and quality scores rose by about 18 percent in that setting; follow-up surveys found higher later job use among the treated group. Graders, not clients, scored the work, and many outputs were lightly edited. Enjoyment rose during the experiment; concern and excitement about AI also rose temporarily. Read the study as evidence that some mid-level writing can change under those conditions. Do not convert it into a promised saving, a staffing cut, or a reason to skip a factual check.
Read a field rollout as one job, not every job. Brynjolfsson, Li and Raymond used staggered access to a suggestion tool for technical customer-support chats. Average productivity, measured as issues resolved per hour, rose; less experienced agents gained more on speed and quality; the most experienced saw small speed gains and small quality declines. Customers were more polite and less likely to ask for a manager in the published account. Agents still chose whether to follow a suggestion. A conversational assistant trained on better workers’ practices is not the same product as a public chatbot, and a support queue is not a courtroom, a clinic or a classroom.
Use the jagged frontier as a check, not a slogan. Dell’Acqua, Ethan Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Karim Lakhani and colleagues assigned 758 consultants, about 7 percent of BCG individual contributors, to no AI, GPT-4, or GPT-4 plus a short prompting overview. On 18 tasks inside the then-current frontier, AI users completed more work faster with higher-rated quality; below-average baseline performers gained more than above-average ones. On one outside-frontier task, AI users were less likely to be correct. The practical move is to notice when a task only looks similar to one the tool can do. If you cannot tell, add a check that does not depend on the model agreeing with itself.
Keep over-trust in view without turning work into homework. Bastani’s GPT-4 tutor study is a high-school math field experiment, not a workplace trial. The workplace-adjacent lesson is that a fluent helper can carry the thinking you will later be asked to stand behind. NIST’s human-AI configuration risk includes over-reliance, automation bias and treating a system as if it were a person. Appendix C of the AI RMF asks organizations to define human roles, including when a person must still be able to overrule the output. At work, that looks like a named reviewer and a later unaided reconstruction of the decision, not a ban on adult tool use.
Ask how an automated workplace decision can be questioned. When AI influences scheduling, performance assessment or another decision about workers, knowing that a tool was used is only a start. Ask what information went in, what the output affected, who is responsible for the final decision and how an error can be raised. NIST’s risk framework includes engagement with affected people and documented feedback about impacts. These questions are a practical way to apply that framework, not a statement that every jurisdiction provides the same legal rights. Employer policy, worker representation and relevant local rules determine the formal route.
Use professional confidentiality as a pattern, not a copy-paste rule. Lawyers have Model Rule 1.6 and Formal Opinion 512; clinicians, accountants, teachers and government staff often have parallel duties from statute, license or contract. The shared pattern is that the input can be the breach. ICO guidance distinguishes the accuracy of personal data about someone from the statistical accuracy of a model, and it keeps minimisation in view. FTC staff have said providers should keep confidentiality and training-data promises. Your sector’s rule controls. When the rule is unclear, keep the material out of a public tool and ask the person who owns the policy.
Let GOVERN, MAP, MEASURE and MANAGE shrink to one task. Tabassi’s AI RMF is written for organizations, including small ones, and is explicitly voluntary. GOVERN asks who is accountable; MAP asks what the system is for and who could be harmed; MEASURE asks how you will notice failure; MANAGE asks what you will do when it fails. The generative-AI profile adds risks that are new or worse for these tools, including confabulation, data privacy, information security, intellectual property and human-AI configuration. Reva Schwartz is a named coauthor of that profile. On an ordinary desk, the miniature version is: I am still the sender; this is a draft of X; I will check Y; if Y is wrong I will not send.
Measure the finished work, including the checking. For a permitted, low-stakes task using public or otherwise approved material, define a satisfactory result before comparing workflows. A meeting-summary check might look for missing decisions, invented commitments and wrong names, as well as the time spent drafting and correcting. Keep the task and standard comparable when judging whether AI helped. This is a practical application of NIST’s call for context-specific measurement, not a validated experiment or a universal benchmark. Faster first output can still mean slower finished work if review and rework grow. Use the result to adjust this task’s workflow rather than generalizing to every task in the job.

related

career
learning
information

watch For Long

Good to know. This is a reading companion about ordinary workplace drafting, checking, privacy, over-trust, and dividing a job into tasks. It is not legal, human-resources, medical, or financial advice, a productivity promise, or a vendor recommendation. Use your employer’s current policy and approved tools when they exist. Keep client, patient, student, employee, and employer-confidential material out of tools you do not control. Safety-critical, licensed, or regulated work still needs the competence and sign-off the role requires. For studying with a chatbot, see Learning and remembering. For checking an online claim or an AI image, see Making sense of online information.
Not advice. Named sources. Honest paraphrase of the finding. Not medical, legal, or financial advice.

sources

Erik Brynjolfsson. Jerry Yang and Akiko Yamazaki Professor and Senior Fellow at Stanford HAI, and director of the Stanford Digital Economy Lab. His contribution here is the customer-support field study of a generative assistant, with uneven gains by experience. It is one workplace, not a universal productivity result.
Danielle Li. David Sarnoff Professor of Management of Technology at MIT Sloan and NBER faculty research fellow. She helps explain how a suggestion tool can spread better workers’ practices to newer agents. The study does not prescribe a staffing model or a public chatbot.
Lindsey Raymond. economist; coauthor of Generative AI at Work while a doctoral researcher at MIT Sloan, later a Microsoft Research postdoctoral researcher and MIT faculty appointee. Her contribution is the workplace rollout evidence. A support-chat assistant is not every generative tool.
Fabrizio Dell’Acqua. researcher of human-AI collaboration; lead author of the BCG GPT-4 field experiment, with appointments moving from Harvard Business School research roles toward Michigan Ross. His jagged-frontier result shows gains inside a capability boundary and worse correctness outside it. It is not a claim about every occupation.
Ethan Mollick. Wharton associate professor of management, co-director of Wharton’s Generative AI Labs, and public educator through One Useful Thing. He translates the jagged frontier for practitioners. Newsletter reach and a popular book are not independent validation of every workplace adaptation.
Hamsa Bastani. Wharton associate professor of operations, information and decisions. Her GPT-4 tutor field experiment is used here only as workplace-adjacent evidence that a fluent helper can become a crutch. It is high-school mathematics, not a ban on adult work tools.
David Autor. Daniel and Gail Rubinfeld Professor at MIT Economics and co-director of the Stone Center on Inequality and Shaping the Future of Work. His task view of automation keeps substitution and complementarity both in the picture. Historical patterns do not forecast one person’s job.
Pawel Gmyrek. ILO senior researcher and coauthor of Generative AI and Jobs. He helps keep exposure at the level of tasks and occupations, with uncertainty about net employment. A global index is not a personal layoff notice.
Janine Berg. ILO senior economist and coauthor of the same jobs analysis. She emphasizes job quality, fair transitions and social dialogue rather than a single automation percentage. Policy discussion is not individualized career advice.
Elham Tabassi. lead author of NIST AI RMF 1.0 while Associate Director for Emerging Technologies at NIST’s Information Technology Laboratory; later NIST chief AI advisor and, after NIST, director of Brookings’ AI and Emerging Technology Initiative. The framework is voluntary and does not certify a particular product.
Arvind Narayanan. Princeton professor of computer science and director of the Center for Information Technology Policy; coauthor of AI Snake Oil. He helps readers separate advertised capability from demonstrated function. Public education about hype is not a test of one workplace tool.
Ifeoma Ajunwa. Asa Griggs Candler Professor of Law at Emory and founding director of its AI and the Future of Work Program. Her work on automated hiring, monitoring and the quantified worker keeps accountability and civil-rights questions visible. It is scholarship and public education, not a ruling in a particular case.
Stijn Broecke. OECD senior economist leading Future of Work research and coauthor of the 2023 employer and worker AI surveys. Those surveys report perceptions in manufacturing and finance in seven countries. They are not causal proof that AI improved every workplace.
Katherine Kellogg. David J. McGrath Jr Professor of Management and Innovation at MIT Sloan, studying how knowledge workers implement AI. She helps translate the jagged frontier and the need to include users in design and training. Implementation research does not guarantee a successful rollout.
Reva Schwartz. NIST researcher and named coauthor of the 2024 Generative AI Profile. That profile names confabulation, data privacy and human-AI configuration among risks unique to or worsened by generative systems. It is voluntary guidance, not a detector you can run on one email.

source Links

NOY-ZHANG-2023

name

Noy S, Zhang W. Experimental evidence on the productivity effects of generative artificial intelligence. Science. 2023;381(6654):187–192. Published July 13, 2023.

kind

evidence

links

label
DOI
href
https://www.science.org/doi/10.1126/science.adh2586
label
PubMed
href
https://pubmed.ncbi.nlm.nih.gov/37440646/
label
CEPR/VoxEU
href
https://cepr.org/voxeu/columns/productivity-effects-generative-artificial-intelligence

BRYNJOLFSSON-QJE-2025

name

Brynjolfsson E, Li D, Raymond L. Generative AI at Work. Quarterly Journal of Economics. 2025;140(2):889–942.

kind

evidence

links

label
DOI
href
https://doi.org/10.1093/qje/qjae044
label
arXiv abstract, v2 6 Nov 2024
href
https://arxiv.org/abs/2304.11771
label
NBER w31161
href
https://www.nber.org/papers/w31161/

DELLACQUA-FRONTIER-2023

name

Dell’Acqua F, McFowland III E, Mollick ER, Lifshitz-Assaf H, Kellogg K, Rajendran S, Krayer L, Candelon F, Lakhani KR. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013; posted September 15, 2023.

kind

evidence

links

label
SSRN
href
https://papers.ssrn.com/abstract=4573321

NIST-AIRMF-2023

name

Tabassi E. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. January 2023.

kind

evidence

links

label
DOI
href
https://doi.org/10.6028/NIST.AI.100-1
label
PDF
href
https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
label
Framework page
href
https://www.nist.gov/itl/ai-risk-management-framework

NIST-GAI-2024

name

Autio C, Schwartz R, Dunietz J, Jain S, Stanley M, Tabassi E, Hall P, Roberts K. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1. July 26, 2024.

kind

evidence

links

label
DOI
href
https://doi.org/10.6028/NIST.AI.600-1
label
NIST publication record
href
https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

OECD-AI-WORK-2023

name

Lane M, Williams M, Broecke S. The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers. OECD Social, Employment and Migration Working Papers No. 288. March 27, 2023.

kind

evidence

links

label
DOI
href
https://doi.org/10.1787/ea0a0fe1-en
label
OECD page
href
https://www.oecd.org/en/publications/the-impact-of-ai-on-the-workplace-main-findings-from-the-oecd-ai-surveys-of-employers-and-workers_ea0a0fe1-en.html

ILO-AI-2023

name

Gmyrek P, Berg J, Bescond D. Generative AI and Jobs: A global analysis of potential effects on job quantity and quality. ILO Working Paper 96. Geneva: ILO; August 21, 2023.

kind

evidence

links

label
Official publication
href
https://www.ilo.org/publications/generative-ai-and-jobs-global-analysis-potential-effects-job-quantity-and
label
DOI
href
https://doi.org/10.54394/FHEM8239

ABA-OPINION-512-2024

name

ABA Standing Committee on Ethics and Professional Responsibility. Formal Opinion 512: Generative Artificial Intelligence Tools. July 29, 2024.

kind

evidence

links

label
ABA news release
href
https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/
label
PDF
href
https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf

FTC-AI-PRIVACY-2024

name

Federal Trade Commission. AI Companies: Uphold Your Privacy and Confidentiality Commitments. Tech at FTC, January 8, 2024.

kind

evidence

links

label
Official post
href
https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/01/ai-companies-uphold-your-privacy-confidentiality-commitments

ICO-AI-DP-2023

name

UK Information Commissioner’s Office. Guidance on AI and data protection. Updated March 15, 2023.

kind

evidence

links

label
Official guidance
href
https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/

BASTANI-LEARNING-2024

name

Bastani H, Bastani O, Sungu A, Ge H, Kabakcı Ö, Mariman R. Generative AI Can Harm Learning. SSRN working paper, posted July 15, 2024.

kind

evidence

links

label
SSRN
href
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4895486
label
DOI
href
https://doi.org/10.2139/ssrn.4895486

AUTOR-JOBS-2015

name

Autor DH. Why Are There Still So Many Jobs? The History and Future of Workplace Automation. Journal of Economic Perspectives. 2015;29(3):3–30.

kind

evidence

links

label
JEP
href
https://www.aeaweb.org/articles?id=10.1257/jep.29.3.3
label
MIT DSpace
href
https://dspace.mit.edu/handle/1721.1/109476

BRYNJOLFSSON-AIAW-2026

name

Erik Brynjolfsson

kind

professional

links

label
Stanford Economics
href
https://economics.stanford.edu/people/erik-brynjolfsson
label
Digital Economy Lab
href
https://digitaleconomy.stanford.edu/person/erik-brynjolfsson/

LI-AIAW-2026

name

Danielle Li

kind

professional

links

label
MIT Sloan directory
href
https://mitsloan.mit.edu/faculty/directory/danielle-li

RAYMOND-AIAW-2026

name

Lindsey Raymond

kind

professional

links

label
DEL publication page
href
https://digitaleconomy.stanford.edu/publication/generative-ai-at-work/

DELLACQUA-AIAW-2026

name

Fabrizio Dell’Acqua

kind

professional

links

label
Personal academic page
href
https://www.fabriziodellacqua.com/

MOLLICK-AIAW-2026

name

Ethan Mollick

kind

professional

links

label
Knowledge at Wharton
href
https://knowledge.wharton.upenn.edu/faculty/ethan-mollick/
label
One Useful Thing
href
https://www.oneusefulthing.org/about

BASTANI-AIAW-2026

name

Hamsa Bastani

kind

professional

links

label
Faculty site
href
https://hamsabastani.github.io/
label
Wharton Executive Education
href
https://executiveeducation.wharton.upenn.edu/faculty/hamsa-bastani/

AUTOR-AIAW-2026

name

David Autor

kind

professional

links

label
MIT Economics
href
https://economics.mit.edu/people/faculty/david-h-autor

GMYREK-AIAW-2026

name

Pawel Gmyrek

kind

professional

links

label
ILO publication
href
https://www.ilo.org/publications/generative-ai-and-jobs-global-analysis-potential-effects-job-quantity-and

BERG-AIAW-2026

name

Janine Berg

kind

professional

links

label
ILO research profile
href
https://researchrepository.ilo.org/esploro/profile/janine_berg

TABASSI-AIAW-2026

name

Elham Tabassi

kind

professional

links

label
NIST AI RMF
href
https://www.nist.gov/itl/ai-risk-management-framework
label
Brookings
href
https://www.brookings.edu/people/elham-tabassi/

NARAYANAN-AIAW-2026

name

Arvind Narayanan

kind

professional

links

label
Princeton page
href
https://www.cs.princeton.edu/~arvindn/

AJUNWA-AIAW-2026

name

Ifeoma Ajunwa

kind

professional

links

label
Emory AI.Humanity directory
href
https://aihumanity.emory.edu/faculty/faculty-hires.html
label
Berkman Klein
href
https://cyber.harvard.edu/people/iajunwa

BROECKE-AIAW-2026

name

Stijn Broecke

kind

professional

links

label
OECD.AI community
href
https://oecd.ai/en/community/stijn-broecke

KELLOGG-AIAW-2026

name

Katherine Kellogg

kind

professional

links

label
MIT Sloan directory
href
https://mitsloan.mit.edu/faculty/directory/kate-kellogg

SCHWARTZ-AIAW-2026

name

Reva Schwartz

kind

professional

links

label
NIST publication record
href
https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

DELLACQUA-FRONTIER-WAVE

name

Dell’Acqua et al.: Navigating the Jagged Technological Frontier, published 2026. Source

kind

evidence

links

label
Source
href
https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf

BRYNJOLFSSON-AI-WAVE

name

Brynjolfsson, Li and Raymond: Generative AI at Work. Source

kind

evidence

links

label
Source
href
https://digitaleconomy.stanford.edu/publication/generative-ai-at-work/

ILO-AI-2025

name

ILO–NASK: One in four jobs potentially exposed to transformation by GenAI. Source

kind

evidence

links

label
Source
href
https://www.ilo.org/resource/news/one-four-jobs-risk-being-transformed-genai-new-ilo%E2%80%93nask-global-index-shows

NIST-MEASURE-WAVE

name

NIST AI Risk Management Framework 1.0. Source

kind

evidence

links

label
Source
href
https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf

source Claims

watch For

ILO-AI-2023
ABA-OPINION-512-2024
NIST-AIRMF-2023

situations.draft

NOY-ZHANG-2023

situations.check

NIST-GAI-2024
DELLACQUA-FRONTIER-2023
MOLLICK-AIAW-2026

situations.private

ABA-OPINION-512-2024
NIST-AIRMF-2023
FTC-AI-PRIVACY-2024
ICO-AI-DP-2023

situations.divide

ILO-AI-2023
AUTOR-JOBS-2015
GMYREK-AIAW-2026
BERG-AIAW-2026
AUTOR-AIAW-2026
ILO-AI-2025

split.established.0

ILO-AI-2023
AUTOR-JOBS-2015
GMYREK-AIAW-2026
BERG-AIAW-2026
ILO-AI-2025

split.established.1

NOY-ZHANG-2023

split.established.2

BRYNJOLFSSON-QJE-2025
BRYNJOLFSSON-AIAW-2026
LI-AIAW-2026
RAYMOND-AIAW-2026
KELLOGG-AIAW-2026

split.established.3

ABA-OPINION-512-2024
ICO-AI-DP-2023
NIST-AIRMF-2023
FTC-AI-PRIVACY-2024

split.established.4

NIST-GAI-2024
NIST-AIRMF-2023
NARAYANAN-AIAW-2026
SCHWARTZ-AIAW-2026

split.established.5

NIST-AIRMF-2023
TABASSI-AIAW-2026

split.disputed.0

OECD-AI-WORK-2023
BROECKE-AIAW-2026

split.disputed.1

DELLACQUA-FRONTIER-2023
BRYNJOLFSSON-QJE-2025
MOLLICK-AIAW-2026
DELLACQUA-AIAW-2026
DELLACQUA-FRONTIER-WAVE
BRYNJOLFSSON-AI-WAVE

split.disputed.2

ABA-OPINION-512-2024
AJUNWA-AIAW-2026

split.disputed.3

ILO-AI-2023
OECD-AI-WORK-2023
AUTOR-JOBS-2015
ILO-AI-2025

split.talk.0

BASTANI-LEARNING-2024
BASTANI-AIAW-2026

split.talk.1

NIST-AIRMF-2023
NARAYANAN-AIAW-2026

split.talk.2

ABA-OPINION-512-2024
FTC-AI-PRIVACY-2024
ICO-AI-DP-2023

split.talk.3

ILO-AI-2023
AUTOR-JOBS-2015
AJUNWA-AIAW-2026

move

NOY-ZHANG-2023
ABA-OPINION-512-2024
NIST-AIRMF-2023

try.0

NOY-ZHANG-2023

try.1

NIST-GAI-2024
DELLACQUA-FRONTIER-2023

try.2

ILO-AI-2023
AUTOR-JOBS-2015
BRYNJOLFSSON-QJE-2025

keep.0

OECD-AI-WORK-2023

keep.1

ABA-OPINION-512-2024
ICO-AI-DP-2023
NIST-AIRMF-2023

keep.2

NIST-AIRMF-2023

keep.3

BASTANI-LEARNING-2024

keep.4

KELLOGG-AIAW-2026

deep.0

NOY-ZHANG-2023

deep.1

BRYNJOLFSSON-QJE-2025
LI-AIAW-2026
RAYMOND-AIAW-2026

deep.2

DELLACQUA-FRONTIER-2023
MOLLICK-AIAW-2026
KELLOGG-AIAW-2026
DELLACQUA-AIAW-2026

deep.3

BASTANI-LEARNING-2024
NIST-AIRMF-2023
NIST-GAI-2024
BASTANI-AIAW-2026

deep.4

AJUNWA-AIAW-2026
NIST-MEASURE-WAVE

deep.5

ABA-OPINION-512-2024
ICO-AI-DP-2023
FTC-AI-PRIVACY-2024

deep.6

NIST-AIRMF-2023
NIST-GAI-2024
TABASSI-AIAW-2026
SCHWARTZ-AIAW-2026

deep.7

ILO-AI-2023
BASTANI-LEARNING-2024
NIST-MEASURE-WAVE
DELLACQUA-FRONTIER-WAVE

claims.0

NOY-ZHANG-2023

claims.1

ILO-AI-2023
AUTOR-JOBS-2015
ILO-AI-2025

claims.2

ABA-OPINION-512-2024
ICO-AI-DP-2023
FTC-AI-PRIVACY-2024

claims.3

NIST-GAI-2024
DELLACQUA-FRONTIER-2023

claims.4

BRYNJOLFSSON-QJE-2025
DELLACQUA-FRONTIER-2023

claims.5

OECD-AI-WORK-2023

claims.6

BASTANI-LEARNING-2024
ILO-AI-2023
AUTOR-JOBS-2015

source Situation Keys

situations.draft
situations.check
situations.private
situations.divide

professional Source Ids

BRYNJOLFSSON-AIAW-2026
LI-AIAW-2026
RAYMOND-AIAW-2026
DELLACQUA-AIAW-2026
MOLLICK-AIAW-2026
BASTANI-AIAW-2026
AUTOR-AIAW-2026
GMYREK-AIAW-2026
BERG-AIAW-2026
TABASSI-AIAW-2026
NARAYANAN-AIAW-2026
AJUNWA-AIAW-2026
BROECKE-AIAW-2026
KELLOGG-AIAW-2026
SCHWARTZ-AIAW-2026

house Pointers

href
/learn/
label
STUDlearn. For burnout, work design, and career experiments, see Career and work.

positive Guidance

status: DONE
reviewed At: 2026-09-12
professional Count: 15
scope: content-and-positive-titles-only

learn Article

article

format version

presentation

complete-card

overview

sections

id
understand-the-assistant
role
explanation
title
What a workplace AI assistant does
intro
A draft, a verified answer and an action taken on your behalf require different kinds of trust.
items
- label
  Give the assistant a defined piece of work
  text
  Generative AI creates content such as text or code from instructions and supplied material. A large language model is one kind of generative system: it learns patterns in language and uses them to produce a response. Your prompt is the request and context you give it. That can help turn a rough idea into an outline, rewrite a paragraph or propose questions. It does not give the system automatic access to your company's actual promises, policies or records. Brynjolfsson, Li and Raymond distinguish the underlying model from the workplace system built around it. Begin by naming the output you want and which facts it is allowed to use.
  source ids
  support-qje-2025
  ncsc-llm-2023
- label
  Use source material to make an answer checkable
  text
  Some assistants can search documents or retrieve passages before generating an answer. Retrieval means finding material; generation means composing the response. These are separate operations: finding the right document does not prove that the summary preserves its meaning. The support assistant studied by Brynjolfsson and colleagues provided suggestions and documentation links, while agents retained discretion over the conversation. At your desk, a useful answer includes a passage or locator you can open, not merely a convincing reference to a handbook. Check that the cited passage supports the particular statement and that the document is the current one.
  source ids
  support-qje-2025
  nist-measure
- label
  Review an action before granting permission to carry it out
  text
  An assistant that drafts an email differs from an agent that can send it. Agentic AI is connected to tools and can take actions toward a goal, potentially without continuous human intervention. The NCSC's May 2026 guidance says organizations should define ownership, access, monitoring and the ability to stop the system before connecting agents to real work. For an ordinary user, the practical distinction is straightforward: a draft can be edited; a sent message, changed record or executed command may already affect someone. Use only an approved arrangement, with the access and review your organization requires. A successful drafting trial does not establish readiness for autonomous action.
  source ids
  ncsc-agents-2026
id
choose-a-work-task
role
application
title
Where it can fit into an ordinary workday
intro
These are practical applications of the evidence, not tested STUDlearn procedures or promised savings.
items
- label
  Start an email with its purpose and audience
  text
  When you know what an email must accomplish but struggle to start, ask for a structure using public facts or fictional placeholders. For example: draft a brief meeting invitation with spaces for the date, location and reply deadline. Noy and Zhang tested occupation-specific writing assignments and found a useful role for assistance with composition. Their tasks did not depend on private organizational knowledge or precise factual accuracy, so the result cannot certify a real client message. Put the correct facts into the draft on an authorized system. Read it for tone, invented commitments and whether the recipient can tell what to do next. You remain responsible for the send.
  source ids
  writing-2023
  writing-author-account
- label
  Compare a summary with the document it represents
  text
  When a long notice or meeting record feels hard to digest, an approved assistant may help organize it. A summary is a shorter account of a source, so its quality depends on preserving decisions, exceptions and uncertainty rather than merely sounding clear. NIST recommends evaluating AI in the context in which it will be used. For a routine summary, compare the output with the original and ask: did it miss a decision, turn a suggestion into a commitment, or attach the wrong person or date? Keep the record available for the reader. If you cannot open the underlying source or competently review it, the summary is not a reliable basis for your decision. Consider a fictional notice: a workshop is planned for Thursday, subject to the room being confirmed, and attendees must reply by Tuesday. An assistant summarizes it as 'The workshop is Thursday; reply by Thursday.' The sentence is fluent, but it has removed a condition and changed a deadline. A source check has to compare those details, not merely reread the summary for good grammar. The reviewer restores 'planned, pending room confirmation,' corrects the reply date and leaves a source locator. If the room is later confirmed, that is a new fact requiring an updated record. Counting the quick first draft as time saved would omit this review and correction work. This invented example shows what responsible review actually checks; it does not establish a particular tool's accuracy or authorize sharing a real internal notice with it.
  source ids
  nist-measure
  nist-gai
- label
  Choose a permitted route before supplying work information
  text
  When the task involves a client file, staff record or unreleased number, decide whether the tool and that particular use are authorized before entering anything. Confidentiality concerns what may be shared, with whom and for what purpose; it is not just a setting labeled private. The NCSC describes shadow AI as use outside an organization's approved systems and processes, which can reduce visibility and control of information. Replace actual details with invented examples only when they no longer reveal the underlying matter. Otherwise use the approved environment or do the work without the tool. An app installed on a work laptop is not, by itself, permission to upload every file.
  source ids
  ncsc-shadow-2026
  ico-minimisation
  aba-512
- label
  Discuss a changing task rather than your whole job at once
  text
  When a headline makes your occupation sound obsolete, identify one task the tool may change and one responsibility that remains. Drafting a report, verifying its figures and explaining its recommendation are different parts of a job. The ILO's 2025 index measures potential exposure of occupational tasks to generative AI; it does not count layoffs. Autor's task framework explains why automation can substitute for some work and complement other work. This does not guarantee employment security. It makes a workplace conversation more useful: what will be automated, what expertise is needed to review it, and what training, staffing or sign-off will support the new process?
  source ids
  ilo-exposure-2025
  autor-2015
id
read-the-workplace-evidence
role
findings
title
What the workplace studies actually found
intro
The studies answer different questions. Their results describe particular people, tasks and systems, rather than a single productivity rate for AI.
items
- label
  Some short writing tasks improved with AI assistance
  text
  Shakked Noy and Whitney Zhang randomly gave ChatGPT access to half of 453 college-educated professionals completing paid, occupation-specific writing tasks in 2023. Participants were managers, marketers, grant writers, consultants, human-resource professionals and data analysts. Average completion time fell about 40 percent and judged quality rose about 18 percent. Lower-baseline writers gained more. The assignments included reports, press releases and tactful emails, and experienced graders assessed the output. The authors emphasize that the tasks favored clear, relatively generic writing and did not require context-specific knowledge or precise factual accuracy. That makes the finding useful for drafting, while leaving a real workplace's fact checking, authorization and audience requirements to be tested separately.
  source ids
  writing-2023
- label
  A useful draft can change how people work without proving long-term competence
  text
  In their public account, Noy and Zhang reported that many participants submitted unedited or lightly edited chatbot output and graders often rated it well. That describes what worked in their experiment; it is not an instruction to forward raw output at work. Participants enjoyed the assisted task more, and both excitement and concern about AI rose temporarily. Those attitude effects faded in follow-up surveys, while treated participants were more likely to report later workplace use. Adoption, enjoyment and a good graded document are different outcomes from sustained competence or a client's satisfaction. Keep that distinction when someone offers a compelling demonstration as proof of lasting value.
  source ids
  writing-author-account
  writing-2023
- label
  Support workers gained unevenly from a specialized suggestion tool
  text
  Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied staggered access to a conversational assistant among 5,172 customer-support agents. The published 2025 study found about 15 percent more issues resolved per hour on average. Less experienced and lower-skilled agents improved speed and quality; the most experienced and highest-skilled agents had small speed gains and small quality declines. Customers were more polite and less likely to request a manager. Agents could edit or ignore suggestions. This was a tool adapted to one firm's technical-support work, rather than an unrestricted public chatbot. A team should therefore examine whose work improves and whether quality changes, rather than assign the average gain to everybody.
  source ids
  support-qje-2025
- label
  Check capability on the task you actually need done
  text
  Dell'Acqua and colleagues call this the jagged technological frontier: AI can help with one task and hurt performance on another that looks equally difficult to a person. Their experiment included 758 BCG consultants, roughly 7 percent of individual contributors, in early stages of their consulting careers. Participants received no AI, GPT-4, or GPT-4 with a prompting overview. Across 18 tasks within the tested capability range, AI users completed 12.2 percent more tasks, worked 25.1 percent faster and produced higher-rated work. Lower-baseline performers gained more. On a deliberately selected task outside that range, the combined AI groups were about 19 percentage points less likely to reach the correct answer. The 2026 journal article preserves the key contrast. It does not locate today's boundary for your task or certify your preferred prompt.
  source ids
  frontier-2026
- label
  Measure completed work rather than the feeling of moving faster
  text
  METR randomized AI access on 246 real tasks completed by 16 experienced open-source developers in familiar repositories during early 2025. AI access made completion take about 19 percent longer, although developers believed it had helped them work faster. This is a small, specific study, not proof that coding assistance always slows work. In February 2026, METR reported that a later experiment's results were unreliable as a measure of current uplift because participation, task selection and concurrent-agent time measurement had changed. The researchers thought gains were likely larger by then, but could not reliably quantify the increase. Both reports support checking your own finished workflow rather than promoting an old result into a timeless claim.
  source ids
  metr-2025
  metr-update-2026
- label
  Exposure estimates describe potential change rather than observed job loss
  text
  The ILO–NASK 2025 index combines worker task assessments, expert input and model predictions to estimate how generative AI could affect occupations. About one in four workers globally are in occupations with some exposure, and 3.3 percent of employment falls in the highest category. Clerical work remains especially exposed; exposure also differs by income group and sex because occupational employment differs. The authors consider transformation more likely than complete replacement because many jobs retain tasks requiring human input. These are estimates of potential, not a count of displaced people, a forecast for a particular employer, or reassurance that no job will disappear. Use them to identify where transition support may be needed.
  source ids
  ilo-exposure-2025
  ilo-brief-2025
id
protect-work-information
role
cautions
title
Handle work information before it enters a tool
items
- label
  Check approval for the task, data and system together
  text
  A company may approve one AI service for one class of information while restricting another use. Before uploading, establish what data the approved account may receive, who can access inputs and outputs, how they are retained, and whether connected services are covered. The NCSC's September 2026 shadow-AI guidance explains how unapproved use can bypass established visibility and controls. If the policy is unclear, bring a concrete example of the proposed task to the policy owner without exposing the actual confidential file. Approval answers whether you may use the tool; it does not answer whether its output is correct.
  source ids
  ncsc-shadow-2026
  ncsc-llm-2023
- label
  Use only the information needed for the purpose
  text
  Data minimisation means using the personal information necessary for a defined purpose, rather than providing everything that might be convenient. The UK's ICO treats this as a continuing requirement when organizations use AI. At a desk, an editorial application is to replace a real person's details with a genuinely fictional example when you only need help with wording or structure. Removing a name alone may leave a person identifiable through role, dates or circumstances. Keep patient and student records, client matters, employee details, privileged material, credentials, restricted code and unreleased figures out of unapproved tools. A full-file upload deserves more scrutiny than a public paragraph. This is privacy literacy, not a jurisdiction-specific compliance determination.
  source ids
  ico-minimisation
  nist-gai
  ncsc-shadow-2026
- label
  Treat confidentiality as a duty attached to the information
  text
  ABA Formal Opinion 512 explains lawyers' duties of competence, confidentiality, communication, supervision and reasonable fees when using generative AI. Its discussion of self-learning tools says informed consent can be necessary before entering information relating to a representation, and generic boilerplate is insufficient. That opinion concerns lawyers and US model ethics rules; it does not impose one consent procedure on every profession. Teachers, clinicians, accountants and government staff must use their own applicable rules and approved processes. The common workplace question is whether this transfer is allowed before it happens. Deleting a conversation later cannot retroactively authorize a disclosure or restore control of a copy already transferred.
  source ids
  aba-512
  ncsc-llm-2023
- label
  Separate a provider's privacy promise from your permission to share
  text
  A promise not to train on customer data addresses one use of information. It does not, on its own, answer retention, access, security or your employer's authorization requirements. FTC staff explain that providers can face liability when they break privacy and confidentiality commitments, including promises about model training. That is guidance directed at providers, rather than a declaration that a particular work file is safe to upload. Likewise, a prompt does not necessarily update a model instantly. Read the current approved terms and workplace policy for the system you actually use. Treat the transfer of restricted information as the decision to resolve before composing a prompt.
  source ids
  ftc-privacy-2024
  ncsc-llm-2023
id
check-the-finished-work
role
application
title
What a useful check looks like
items
- label
  Compare claims with evidence outside the conversation
  text
  NIST calls confidently false generated content confabulation, often called hallucination. Automation bias means giving an automated suggestion more weight than its evidence deserves. These concepts explain why a polished answer can pass a casual read while still containing the wrong figure or unsupported conclusion. For a factual draft, compare names, dates, amounts and citations with the original record. A citation should exist, address the question and support the sentence. For a calculation, use the authoritative data and an appropriate independent calculation. Asking the same assistant whether it is sure can identify a question to investigate, but its assurance is not that investigation.
  source ids
  nist-gai
  nist-measure
  frontier-2026
- label
  Give the reviewer evidence and authority to change the result
  text
  Human review has value when someone has the knowledge, time and source access to identify an error and can reject or correct the output. In a 2026 MIT account of research on 72 BCG consultants, researchers observed GPT-4 defending recommendations with increasingly persuasive language when users challenged it. That analysis describes interactions in one task, not intentional deceit or every model's behavior. Its practical relevance is that remaining inside a persuasive chat may be a weak check. Move to the original records and record the unresolved point. A named reviewer who merely clicks approval without that comparison has not established reliability.
  source ids
  mit-validation-2026
  nist-govern
- label
  Match review effort to the consequence of an error
  text
  NIST frames risk in terms of both the chance of an event and the severity of its consequences. A wrong heading in a private outline and a wrong number in a client decision have different stakes. Before reviewing, describe the specific failure: an invented deadline, missing exception, wrong recipient or unsupported recommendation. Choose a check capable of finding it, and involve the accountable specialist when the matter needs expertise you lack. Safety-critical, licensed or regulated work retains its required competence and sign-off. A generic 'human checked' label cannot replace that process. If an essential claim remains unverified, keep the output as a draft or use the existing authorized workflow.
  source ids
  nist-rmf
  aba-512
  nist-manage
- label
  Test the complete workflow against a defined standard
  text
  Choose the success criteria before a permitted comparison. For a meeting summary, that might mean retaining all decisions, owners and deadlines, preserving uncertainty, and adding no unsupported commitments. Count time spent preparing the input, generating, reviewing and correcting, not just time until text appears. Compare tasks of similar difficulty against the same quality standard. Danielle Li's public guidance stresses defining success; MIT's 2026 workplace synthesis adds evidence before scaling and attention to rework. This small comparison is a practical adaptation, not a controlled trial. One good outcome supports another cautious test of that task, rather than an all-job productivity percentage.
  source ids
  mit-define-success-2025
  mit-levers-2026
  nist-measure
  metr-2025
- label
  Give the task an owner, a purpose, a check and a fallback
  text
  NIST's voluntary AI Risk Management Framework organizes risk work into four connected functions: GOVERN sets responsibilities and rules; MAP describes the use, affected people and possible harms; MEASURE evaluates performance and risk; MANAGE responds to what is found. They are not an ordered checklist or a certificate of safety. Its trustworthiness characteristics cover validity and reliability, safety, security and resilience, accountability and transparency, explainability and interpretability, privacy, and fairness with harmful bias managed. A desk-level application is to write who owns the result, what the assistant may do, which failure your check can catch and what happens if it fails. The generative-AI profile adds detail on confabulation, privacy, information security, intellectual property and human over-reliance. Correct wording is only one part of responsible use.
  source ids
  nist-rmf
  nist-core
  nist-gai
id
understand-adoption-choices
role
perspectives
title
Where workplace decisions remain unsettled
items
- label
  Ask what training and consultation will change in practice
  text
  OECD surveys of 5,334 workers and 2,053 firms in manufacturing and finance across seven countries found generally positive reported effects alongside concern about job loss. Training and worker consultation were associated with better reported outcomes. Association means the factors occurred together; it does not establish that training caused the improvement. Organizations with good management may differ in other ways. The 2023 report concerns surveys conducted in early 2022, before ChatGPT's public release; it is not a survey of today's generative tools. The genuine unresolved question is which training and consultation arrangements work best for which jobs. Ask who can demonstrate a useful review, who workers can report problems to, and whether staff feedback changes deployment. Optimism and worry both deserve a specific account of what happened.
  source ids
  oecd-surveys-2023
  oecd-survey-design
  professional-kellogg
- label
  Match assistance and review to the worker's experience
  text
  The writing, support and consultant studies often found larger gains for lower-baseline workers within the tasks where assistance worked. That is a shared pattern, rather than opposing findings about novices. Yet the consultant study also demonstrated a correctness drop outside the tested capability range. 'Lower skill' there was relative to other highly trained consultants, not an entry-level worker in every field. A team can reasonably debate which tasks need expert review, how much training new staff require and whether experts should use different workflows. Narrowing an output-quality gap does not prove that professional knowledge, exception handling or supervisory competence has become unnecessary.
  source ids
  writing-2023
  support-qje-2025
  frontier-2026
- label
  Use the disclosure rules that govern the work
  text
  Whether AI assistance must be named depends on the employer, client agreement, professional requirements and circumstances. ABA Opinion 512 describes situations requiring communication or informed consent and situations where routine assistance need not be disclosed; it is not a universal email-label rule. A useful internal handoff can say what was checked, who reviewed it and what remains uncertain, while following any required AI disclosure. Separately, disclosure of a tool used to evaluate a worker is a different issue from disclosure that you used it to draft a paragraph. Ifeoma Ajunwa's workplace-technology scholarship helps keep that distinction visible. Transparency should help the affected person understand the decision and the route for questioning it.
  source ids
  aba-512
  professional-ajunwa
  nist-govern
- label
  Read employment forecasts alongside the choices that shape jobs
  text
  Task exposure, employer adoption and observed employment are different measurements. A forecast should state its date, geography, occupations and assumptions about demand and deployment. Autor explains that technology can substitute for labor while also creating complementary work and changing demand. This does not rule out displacement. In their April 2026 synthesis of early experiments at more than 20 companies, Ben Armstrong and Julie Shah describe substantial variation in how organizations turn AI into supervisory work and whether that work is engaging or repetitive. It is qualitative organizational evidence, not a causal estimate of net jobs. Training, staffing, workflow design and how saved time is used remain real decisions.
  source ids
  autor-2015
  ilo-exposure-2025
  mit-work-design-2026
id
keep-the-skill-you-own
role
application
title
Keep enough skill to supervise the work
items
- label
  Separate assisted performance from what you can do later
  text
  Hamsa Bastani and colleagues tested two GPT-4-based tutors in high-school mathematics: an open-ended helper and a tutor with teacher-informed safeguards. Both improved assisted practice performance, but the open-ended group later scored worse than controls on unaided exams. The safeguarded tutor largely prevented that penalty without demonstrating a later learning gain. The published 2025 study is classroom evidence, not a trial of adult workplace use. It makes a useful question visible: did the tool help you learn a method or simply produce an answer? When the job still requires judgment, close the chat and explain the calculation, exception or recommendation using the source material. Use Learning and remembering for the fuller study-skills question.
  source ids
  bastani-pnas-2025
- label
  Use your knowledge to direct and evaluate the assistant
  text
  A 2025 CHI survey by Hao-Ping Lee, Advait Sarkar, Lev Tankelevitch and colleagues collected 936 examples from 319 knowledge workers. Higher confidence in generative AI was associated with less reported critical-thinking effort, while greater confidence in one's own ability was associated with more. The researchers described a shift toward verification, integrating responses and overseeing tasks. This was self-reported behavior, not an objective test proving that AI destroys thinking skills. The support-agent study also found evidence consistent with learning when assistance was temporarily unavailable. Outcomes can differ with design and use. Maintain chances to practice, explain and challenge the work you remain accountable for, rather than treating either universal learning or universal deskilling as settled.
  source ids
  critical-thinking-2025
  support-qje-2025
id
maintain-a-useful-process
role
application
title
Keep the process useful over time
items
- label
  Keep the current policy and an accountable contact within reach
  text
  Save the current workplace AI policy or the name of the person responsible for the approved service. Check whether the permission covers your specific task and class of information when either changes. NIST's GOVERN guidance connects AI responsibilities to existing organizational rules, while the NCSC explains the risks when informal use escapes those processes. An unclear policy is a reason to seek a concrete answer before supplying restricted material. It does not require abandoning a useful public-material experiment that is already permitted. Policy gives you a boundary; source comparison gives you evidence about the result.
  source ids
  nist-govern
  ncsc-shadow-2026
- label
  Keep sensitive facts on the system authorized to hold them
  text
  Write a reusable prompt around the task's structure rather than real client names, personal histories or unreleased numbers. Fill those facts into the result only in an authorized place, after checking that the wording still fits. ICO minimisation guidance provides the principle of limiting personal data to what the purpose requires. The desk-level example is an editorial adaptation: it is useful when generic structure is enough, but unsuitable if the fictionalized input still reveals a person or the real task cannot be reviewed without restricted information. In that case, use the approved data-handling route or the ordinary workflow.
  source ids
  ico-minimisation
  aba-512
- label
  Keep a source locator with factual work
  text
  Store the approved source location, date or document version beside a factual draft so someone can repeat the comparison later. A sentence such as 'the notice dated October 1, section 3' is useful only if it identifies a real record the reviewer can access. Where records conflict, state what differs and seek the responsible person's resolution; a fluent generated paragraph should not silently choose. NIST's measurement guidance calls for context and documented evaluation. The ICO also distinguishes a model's average predictive accuracy from accurate information about a particular person. If a record contains a prediction about someone, identify it as an inference rather than an established fact. Neither a general accuracy score nor a tidy citation resolves a specific contradictory record.
  source ids
  nist-measure
  ico-accuracy
- label
  Be able to explain the decision you approve
  text
  Before approving work that requires your expertise, explain its reasoning with the authoritative source open and the chatbot closed. Identify the evidence, the exception and what would change your conclusion. This is a practical accountability check, rather than an exam score or a validated anti-deskilling intervention. Bastani's study shows why assisted performance and unaided understanding need separate attention, while the CHI survey suggests checking and task oversight may become a larger part of knowledge work. If you cannot reconstruct a material decision, obtain competent review or do that part yourself before relying on the output.
  source ids
  bastani-pnas-2025
  critical-thinking-2025
  nist-core
- label
  Keep a small permitted record and revisit recurring misses
  text
  Where workplace rules allow it, note the task, approved tool, check performed and one useful correction. Do not create a private archive of confidential prompts or copy restricted files into a personal log. NIST's MANAGE guidance supports monitoring problems and responding to them, and Kellogg's implementation work emphasizes workers' involvement and training. Review recurring misses when the task, source or tool changes. If assistance repeatedly invents commitments or consumes more correction time than it saves, narrow its role or use the established workflow. When AI influences scheduling, hiring or performance assessment, ask what information went in, which output affected the decision, who is responsible for the final decision and how an error can be raised. These apply NIST's feedback principle; employer policy, worker representation and local rules determine the formal route, rather than a uniform legal right assumed here.
  source ids
  nist-manage
  professional-kellogg
  professional-ajunwa
  nist-govern
id
check-common-claims
role
findings
title
Check the claims people repeat
intro
An established finding stays bounded to what its study measured. Read the claim and its evidence together before extending it to your work.
items
- label
  Some writing gains are established within a defined experiment
  text
  Claim: ChatGPT-style tools can change time and quality on some mid-level writing tasks in a bounded experiment. Verdict: established. Noy and Zhang measured shorter completion times and higher grades. This does not establish a client-ready result or a guaranteed saving for your own job.
  source ids
  writing-2023
- label
  Task exposure leaves an occupation's future open
  text
  Claim: Generative AI will replace every job this year. Verdict: not. The ILO's index measures potential task exposure. It does not predict universal replacement, count layoffs or settle demand and adoption. A forecast and an observed employment change require their own evidence.
  source ids
  ilo-exposure-2025
  autor-2015
- label
  Resolve permission before sharing confidential material
  text
  Claim: Confidential client or employer material is safe in a public chatbot if I delete the chat. Verdict: not. Deletion does not retroactively authorize the transfer. Applicable confidentiality duties, approved terms and workplace rules govern the information before it enters the tool.
  source ids
  aba-512
  ftc-privacy-2024
  ncsc-shadow-2026
- label
  Fluent wording still needs a factual check
  text
  Claim: If the chatbot wrote it fluently, it is ready to send. Verdict: not. Fluent false content is a recognized risk. Review against the original evidence and the task's required quality standard before relying on it.
  source ids
  nist-gai
  frontier-2026
- label
  Evaluate gains for the particular worker and task
  text
  Claim: A workplace AI assistant raises every worker's performance equally. Verdict: not. The support study found different effects by experience and skill, and the consultant study found different effects by task. An average is not an individual guarantee.
  source ids
  support-qje-2025
  frontier-2026
- label
  Treat the training association as evidence with a causal limit
  text
  Claim: Training and consulting workers about AI is linked with better reported outcomes. Verdict: established for the OECD survey association. The stronger claim that training causes improvement everywhere remains unproven and open to debate. Association and universal effectiveness are different propositions.
  source ids
  oecd-surveys-2023
- label
  Keep the knowledge needed to explain and supervise your work
  text
  Claim: Using AI at work means I no longer need to know the material. Verdict: talk. Producing a draft does not demonstrate the competence needed to check, explain or approve it. The workplace still determines which responsibilities and expertise your role requires.
  source ids
  critical-thinking-2025
  bastani-pnas-2025
  nist-core
id
try-one-permitted-task
role
next_steps
title
Try one small, permitted task
intro
Keep the first move small enough that you can judge it yourself.
items
- label
  Name the task, your check and the human step that remains
  text
  Choose one low-stakes task you are already allowed to do with AI, such as structuring a meeting invitation from public or fictional information. Write three lines: the assistant will propose the structure; I will check it against the request; I will supply the real facts, choose the tone and decide whether to send. These lines define the assistant's job and the decisions that remain yours. If the actual task needs restricted material, a consequential decision or expertise you lack, use the existing approved process instead of widening this experiment.
  source ids
  autor-2015
  nist-core
  ncsc-agents-2026
- label
  Try a structure with genuine placeholders
  text
  Ask for a short outline with slots for Date, Location and Reply deadline, without entering actual restricted facts. Specify the audience and desired outcome, then compare the proposed structure with the original request. Keep wording that helps, remove any invented promise, and fill real details only on the system authorized to hold them. The writing experiment supports exploring composition help, while minimisation guidance supports keeping unnecessary personal data out. This exact exercise is an editorial example, not the researchers' protocol.
  source ids
  writing-author-account
  ico-minimisation
- label
  Try one source comparison before relying on the result
  text
  Select one generated claim, number or commitment and open the record it should reflect. Mark it matches, wrong or not in the source. Correct or remove a miss, and repeat the review for every material fact before an actual send. This first comparison teaches the method; checking one sentence does not certify a whole document. Record how much review and correction the task needed, without saving secrets. A useful result is a finished piece of work you can explain and stand behind, including the checking time.
  source ids
  nist-measure
  nist-manage
  mit-define-success-2025
id
use-the-right-boundaries
role
cautions
title
Use this with the rules for your work
items
- label
  Use workplace guidance and qualified sign-off for consequential work
  text
  This is a reading companion about ordinary workplace drafting, checking, privacy, over-trust, and dividing a job into tasks. It is not legal, human-resources, medical, or financial advice, a productivity promise, or a vendor recommendation. Use your employer’s current policy and approved tools when they exist. Keep client, patient, student, employee, and employer-confidential material out of tools you do not control. Safety-critical, licensed, or regulated work still needs the competence and sign-off the role requires. For studying with a chatbot, see Learning and remembering. For checking an online claim or an AI image, see Making sense of online information.
  source ids
  nist-core
  aba-512
  ncsc-shadow-2026
- label
  Read named findings as a companion to your judgment
  text
  Named sources. Honest paraphrase of the finding. Not medical, legal, or financial advice. Career and work covers the wider career and work-design question; this topic concentrates on using a workplace assistant. Learning and remembering covers study and memory. Making sense of online information covers evaluating online claims and AI images. Those topics complement this one without replacing the policy, evidence or competence your actual work requires.
  source ids
  nist-core
  autor-2015

related

career
learning
information

sources

id
writing-2023
title
Experimental evidence on the productivity effects of generative artificial intelligence — Noy and Zhang, Science (2023), author-hosted manuscript
url
https://shakkednoy.com/Noy%20Zhang%20NBER%20SI.pdf
kind
study
accessed at
2026-10-01
supports
Opened author-linked manuscript: abstract, experimental design, results and limitations. Verified 453 professionals, randomized access, approximately 40% less time and 18% higher graded quality, lower-baseline gains, attitude changes and follow-up adoption.
limitations
Paid short writing tasks using 2023 ChatGPT; no contextual knowledge or precise factual accuracy requirement. Grader scores are not client outcomes. Science publisher text was unavailable; this author's manuscript corresponds to the 453-person version. An older MIT preprint with 444 participants was excluded from numerical synthesis.
id
writing-author-account
title
The productivity effects of generative artificial intelligence — Noy and Zhang, VoxEU (2023)
url
https://cepr.org/voxeu/columns/productivity-effects-generative-artificial-intelligence
kind
professional
accessed at
2026-10-01
supports
Opened the authors' public account of writing-task design, lightly edited output, enjoyment and later reported workplace use.
limitations
An explanation of the same experiment, not independent replication or advice to send unchecked work. STUDlearn's placeholder exercise is an editorial application.
id
support-qje-2025
title
Generative AI at Work — Brynjolfsson, Li and Raymond, Quarterly Journal of Economics 140(2), 889–942 (2025)
url
https://academic.oup.com/qje/article/140/2/889/7990658
kind
study
accessed at
2026-10-01
supports
Opened published article including abstract, setting, results and learning analysis. Verified 5,172 agents, staggered access, 15% average issues-resolved-per-hour gain, heterogeneous experience/skill effects, customer interaction changes and discretion to ignore suggestions.
limitations
One firm's adapted technical-support assistant; field rollout differs from randomized general-office adoption. Evidence consistent with learning does not certify all unaided competence. Published 15% replaces older working-paper averages.
id
frontier-2026
title
Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua and colleagues, Organization Science (2026)
url
https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf
kind
study
accessed at
2026-10-01
supports
Opened final journal article on publisher and HBS-hosted PDF: design and results for 758 BCG consultants; 18 in-frontier tasks; 12.2% more tasks, 25.1% faster, higher-rated quality and larger lower-baseline gains. Checked outside-frontier group correctness supporting approximately 19 percentage points lower correctness across AI conditions.
limitations
GPT-4 in a specified consultant experiment, not a map of current capability. Final article published online March 11, 2026, DOI 10.1287/orsc.2025.21838. Do not combine its estimates with the 2023 working paper's 40% quality headline or treat percentage-point correctness difference as a universal relative-percent effect.
id
metr-2025
title
Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METR
url
https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
kind
study
accessed at
2026-10-01
supports
Opened research organization's study report: randomized AI access for 16 experienced developers and 246 real repository tasks; 19% slower completion and perceived speedup.
limitations
Small, selected sample, familiar complex repositories and early-2025 tools. Organization's report was reviewed rather than an independent replication. Not all coding or current models.
id
metr-update-2026
title
Update on AI developer productivity — METR (February 24, 2026)
url
https://metr.org/blog/2026-02-24-uplift-update/
kind
study
accessed at
2026-10-01
supports
Opened follow-up report and its discussion of participation/task selection, changed compensation and time measurement with concurrent agents. Authors consider greater uplift likely but say later estimates are unreliable for determining its size.
limitations
Not a clean measured reversal of the 2025 result. Selection and measurement changes prevent a strong contemporary productivity estimate.
id
ilo-exposure-2025
title
Generative AI and Jobs: A Refined Global Index of Occupational Exposure — ILO Working Paper 140 (2025)
url
https://brasil.un.org/sites/default/files/2025-05/OIT-NASK-IAGen_WP140_web.pdf
kind
study
accessed at
2026-10-01
supports
Opened official UN-hosted report, author biographies, abstract and relevant methods/results. Verified task-based worker/expert/model assessment, one in four workers with some occupational exposure, 3.3% highest exposure and transformation-versus-replacement interpretation.
limitations
Modeled potential exposure rather than observed job loss or employer-specific forecast. Employment distributions and the task-rating procedure shape estimates. Relevant portions reviewed, not every page. ILO publication landing page was also opened.
id
ilo-brief-2025
title
One in four jobs at risk of being transformed by GenAI, new ILO–NASK Global Index shows — ILO
url
https://www.ilo.org/resource/news/one-four-jobs-risk-being-transformed-genai-new-ilo%E2%80%93nask-global-index-shows
kind
official
accessed at
2026-10-01
supports
Opened official explanation of the refined 2025 index, variation by occupational composition, income group and sex, and potential transformation.
limitations
Public explanation of the same index, not an additional employment study. Risk of transformation is not a count of layoffs.
id
autor-2015
title
Why Are There Still So Many Jobs? The History and Future of Workplace Automation — David Autor, Journal of Economic Perspectives 29(3), 3–30 (2015)
url
https://www.aeaweb.org/articles?id=10.1257/jep.29.3.3
kind
review
accessed at
2026-10-01
supports
Opened official abstract explaining substitution, complementarity and changing demand as reasons task automation does not directly determine total employment.
limitations
Historical conceptual synthesis, not a 2026 generative-AI employment forecast. Full paper was not reviewed in this pass; application to a three-line task card is editorial.
id
nist-rmf
title
Artificial Intelligence Risk Management Framework (AI RMF 1.0) — NIST AI 100-1 (2023)
url
https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
kind
guideline
accessed at
2026-10-01
supports
Opened framework and relevant scope, trustworthiness, risk, Core and human-role passages. Supports voluntary nature, likelihood/severity, seven trustworthiness characteristics and four connected functions.
limitations
Voluntary framework, not binding law, a certification, or a validated personal checklist. Framework concepts do not establish an organization's compliance.
id
nist-gai
title
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST AI 600-1 (2024)
url
https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
kind
guideline
accessed at
2026-10-01
supports
Opened profile and relevant risk definitions: confabulation, data privacy, information security, intellectual property and human-AI configuration including automation bias and over-reliance. Opened official publication record to verify Reva Schwartz as coauthor.
limitations
Cross-sector risk guidance rather than a prediction that every system always produces these harms. Suggested desk checks are editorial adaptations. Authorship alone does not verify a coauthor's current employment.
id
nist-core
title
AI RMF Core — NIST AI Resource Center
url
https://airc.nist.gov/airmf-resources/airmf/5-sec-core/
kind
guideline
accessed at
2026-10-01
supports
Opened Core: GOVERN, MAP, MEASURE, MANAGE, human oversight, responsibilities, training, affected-party feedback and the statement that functions are not a required ordered checklist.
limitations
Organizational voluntary guidance. Naming a human does not establish meaningful review or a uniform legal right of appeal.
id
nist-govern
title
AI RMF Playbook — GOVERN, NIST
url
https://airc.nist.gov/airmf-resources/playbook/govern/
kind
guideline
accessed at
2026-10-01
supports
Opened relevant policies, responsibility, accountability, training, feedback and oversight suggestions; reviewers need ability to challenge or override results.
limitations
Suggested actions adapted to context, not all mandatory steps or workplace legal requirements. No claim that a policy alone establishes correct output.
id
nist-measure
title
AI RMF Playbook — MEASURE, NIST
url
https://airc.nist.gov/airmf-resources/playbook/measure/
kind
guideline
accessed at
2026-10-01
supports
Opened relevant deployment-context, validity/reliability, generalization-limit, interpretation and ongoing-feedback passages, including MEASURE 2.3, 2.5 and 2.9.
limitations
Framework guidance, not a controlled demonstration of the exact STUDlearn summary or source-comparison exercises.
id
nist-manage
title
AI RMF Playbook — MANAGE, NIST
url
https://airc.nist.gov/airmf-resources/playbook/manage/
kind
guideline
accessed at
2026-10-01
supports
Opened relevant risk-response, monitoring, error/incident communication, appeal and override passages including MANAGE 4.1 and 4.3.
limitations
Context-specific suggested actions; does not create universal worker rights or prescribe a two-minute personal review.
id
aba-512
title
Formal Opinion 512: Generative Artificial Intelligence Tools — American Bar Association (July 29, 2024)
url
https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
kind
guideline
accessed at
2026-10-01
supports
Opened opinion, relevant competence, confidentiality, informed consent, communication, supervision and fees passages. Reviewed limits on self-learning-tool inputs and insufficient boilerplate consent, and context-dependent client disclosure.
limitations
US lawyers and ABA Model Rules; applicability depends on jurisdiction and circumstances. Not legal advice, a general employee consent rule, or a claim that every prompt requires public disclosure.
id
ftc-privacy-2024
title
AI Companies: Uphold Your Privacy and Confidentiality Commitments — FTC Office of Technology (January 2024)
url
https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/01/ai-companies-uphold-your-privacy-confidentiality-commitments
kind
official
accessed at
2026-10-01
supports
Opened staff post explaining potential liability for providers breaking privacy/confidentiality commitments, including promises not to train on customer data.
limitations
Provider-facing staff guidance, not permission for an employee to upload a file. Current individual provider terms were not audited. No-training and no-transfer are different issues.
id
ico-minimisation
title
How should we assess security and data minimisation in AI? — UK Information Commissioner's Office
url
https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/how-should-we-assess-security-and-data-minimisation-in-ai/
kind
guideline
accessed at
2026-10-01
supports
Opened relevant security and minimisation passages: necessary personal data for a purpose, minimising processing throughout development and deployment.
limitations
UK data-protection context; site states guidance is under review following legal changes. General privacy literacy only; replacing a name alone is not a compliance determination or assured anonymisation.
id
ico-accuracy
title
What do we need to know about accuracy and statistical accuracy? — UK Information Commissioner's Office
url
https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/what-do-we-need-to-know-about-accuracy-and-statistical-accuracy/
kind
guideline
accessed at
2026-10-01
supports
Opened distinction between accuracy of personal data and statistical accuracy of models; relevant guidance on marking predictions/inferences rather than presenting them as facts.
limitations
UK legal/regulatory context, under review. Does not determine whether one workplace record is lawful or prescribe a universal record format.
id
ncsc-llm-2023
title
ChatGPT and large language models: what's the risk? — UK National Cyber Security Centre (2023)
url
https://www.ncsc.gov.uk/blog-post/chatgpt-and-large-language-models-whats-the-risk
kind
official
accessed at
2026-10-01
supports
Opened technical explanation of LLMs, prompts, operator access and confidential-input risk. Supports distinction between entering information and a model instantly learning it.
limitations
Dated technical guidance, not verification of current vendor retention policies. Use current approved terms for the particular service. No blanket assertion that every deployment has identical data handling.
id
ncsc-shadow-2026
title
The hidden risks of shadow AI — UK National Cyber Security Centre (September 7, 2026)
url
https://www.ncsc.gov.uk/blogs/the-hidden-risks-of-shadow-ai
kind
official
accessed at
2026-10-01
supports
Opened official explanation of unapproved workplace AI use, diminished organizational visibility and information control, and the importance of usable approved routes and policy.
limitations
Cybersecurity guidance rather than a jurisdiction-specific employment rule. Unopened third-party prevalence statistics in the post were not used.
id
ncsc-agents-2026
title
Thinking carefully before adopting agentic AI — UK National Cyber Security Centre (May 15, 2026)
url
https://www.ncsc.gov.uk/blogs/thinking-carefully-before-adopting-agentic-ai
kind
official
accessed at
2026-10-01
supports
Opened agentic-AI explanation and recommendations on bounded actions/access, ownership, monitoring, incident response, stopping and starting with lower-stakes uses.
limitations
Organizational cyber guidance, not a current vendor setup guide or an assurance that human approval prevents every harm.
id
oecd-surveys-2023
title
The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers — Lane, Williams and Broecke (2023)
url
https://www.oecd.org/en/publications/the-impact-of-ai-on-the-workplace-main-findings-from-the-oecd-ai-surveys-of-employers-and-workers_ea0a0fe1-en.html
kind
study
accessed at
2026-10-01
supports
Opened official publication overview and report; verified manufacturing/finance, seven countries, positive reported outcomes with job-loss concerns, and training/consultation associations. Sample/timing additionally verified on current OECD survey project page.
limitations
Early-2022 perception surveys, not a causal evaluation, a study of current ChatGPT adoption or a survey of all sectors. Relevant report sections were reviewed, not every questionnaire and table. Association does not establish a universal training effect.
id
oecd-survey-design
title
OECD AI surveys of employers and workers — project overview
url
https://www.oecd.org/en/about/projects/aisurveysofemployersandworkers.html
kind
official
accessed at
2026-10-01
supports
Opened About text verifying early-2022 sample of 5,334 workers and 2,053 firms in Austria, Canada, France, Germany, Ireland, UK and US. States surveys repeated in 2026 with wider coverage.
limitations
2026 repetition is not treated as published outcome evidence; this article reports the 2023 publication of 2022 data. Project overview is not an independent study.
id
critical-thinking-2025
title
The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers — Lee, Sarkar, Tankelevitch and colleagues, CHI (2025)
url
https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/
kind
study
accessed at
2026-10-01
supports
Opened author institution's publication record and abstract: 319 workers, 936 examples, confidence associations and shift toward verification, integration and task stewardship; DOI 10.1145/3706598.3713778.
limitations
Self-reports and associations, not an objective causal test of skill loss. Detailed full-paper analyses were not reviewed. Microsoft-employed authors' institutional affiliation is disclosed in panel.
id
bastani-pnas-2025
title
Generative AI without guardrails can harm learning: Evidence from high school mathematics — Bastani and colleagues, PNAS 122(26), e2422633122 (2025)
url
https://www.pnas.org/doi/10.1073/pnas.2422633122
kind
study
accessed at
2026-10-01
supports
Opened published article methods/results: field experiment in Turkish high-school math, GPT-4-based unrestricted and safeguarded tutors, assisted practice improvements, lower later unaided performance with unrestricted helper and mitigation with safeguarded design.
limitations
Classroom mathematics, not adult workplace evidence. Mitigating the unaided penalty is not demonstrating a later learning gain. Use published version rather than imply that the earlier working paper remains the latest evidence.
id
mit-validation-2026
title
How generative AI 'persuasion bombs' users — and how to fight back — MIT Sloan (April 28, 2026)
url
https://mitsloan.mit.edu/ideas-made-to-matter/how-generative-ai-persuasion-bombs-users-and-how-to-fight-back
kind
professional
accessed at
2026-10-01
supports
Opened institutional research explanation of qualitative interaction analysis involving 72 BCG consultants, GPT-4 and 4,339 prompts; discusses increasingly persuasive responses to challenge and verification outside the chat.
limitations
The institutional account was reviewed; the full working paper was unavailable. It describes one experimental setting, not intentional deceit, every model or a universal rate.
id
mit-define-success-2025
title
How to use generative AI to augment your workforce — MIT Sloan (February 24, 2025)
url
https://mitsloan.mit.edu/ideas-made-to-matter/how-to-use-generative-ai-to-augment-your-workforce
kind
professional
accessed at
2026-10-01
supports
Opened Danielle Li's public implementation guidance on defining success, data/infrastructure, workforce effects and evaluating augmentation.
limitations
Professional explanation; not randomized validation of STUDlearn's small workflow comparison or guaranteed return on a tool.
id
mit-levers-2026
title
10 levers shaping whether generative AI truly improves worker performance — MIT Sloan (September 3, 2026)
url
https://mitsloan.mit.edu/ideas-made-to-matter/10-levers-shaping-generative-ai-truly-improves-worker-performance
kind
professional
accessed at
2026-10-01
supports
Opened recent institutional synthesis stressing context, worker judgment, evidence before scaling, knowing when to trust output and costs of rework.
limitations
Professional synthesis, not a single causal estimate for every lever or permission for autonomous work.
id
mit-work-design-2026
title
Humans in the Loop: The Evolution of Work in Early Experiments with Generative AI — Armstrong and Shah, MIT Industrial Performance Center (April 2026)
url
https://ipc.mit.edu/wp-content/uploads/2026/04/Humans_in_the_Loop_full_r01M.pdf
kind
study
accessed at
2026-10-01
supports
Opened report executive summary, methods and relevant findings on early experiments at more than 20 companies in four industry groups, variation in supervisory work and job quality.
limitations
Qualitative interviews/cases, not causal net-employment estimates or representative universal outcomes. Relevant portions reviewed rather than every detail.
id
professional-kellogg
title
Kate Kellogg — MIT Sloan faculty directory
url
https://mitsloan.mit.edu/faculty/directory/kate-kellogg
kind
professional
accessed at
2026-10-01
supports
Opened faculty biography verifying management/innovation role and public organizational research on technology, work practices and implementation; consulted related official MIT explainers listed above.
limitations
Credential and research-context verification, not empirical proof that a specific training or logging scheme will work.
id
professional-ajunwa
title
Ifeoma Ajunwa — Emory Law faculty profile
url
https://law.emory.edu/faculty/faculty-profiles/ajunwa-profile.html
kind
professional
accessed at
2026-10-01
supports
Opened biography identifying Asa Griggs Candler Professor of Law, AI/Future of Work initiative leadership and scholarship on quantified workers, employment technology, privacy and discrimination.
limitations
Professional context distinguishing worker assessment from assistance with one's own draft. Profile is not jurisdiction-specific legal advice or a reviewed causal effect of monitoring.
id
question-speed-perception
title
ExperiencedDevs discussion of the early-2025 developer productivity study
url
https://www.reddit.com/r/ExperiencedDevs/comments/1lwk503/study_experienced_devs_think_they_are_24_faster/
kind
question_signal
accessed at
2026-10-01
supports
Opened discussion to identify a reader question: why can assistance feel faster while measured completion is slower? The answer uses METR's opened reports.
limitations
Anonymous anecdotes and thread popularity are not evidence of productivity, professional consensus or a validated workflow. No factual guidance relies solely on this discussion.
id
question-thinking-headline
title
Futurology discussion of a headline about AI and thinking skills
url
https://www.reddit.com/r/Futurology/comments/1iq1gl2/microsoft_study_finds_relying_on_ai_kills_your/
kind
question_signal
accessed at
2026-10-01
supports
Opened discussion as a signal of confusion between a confidence/self-report association and a claim that AI destroys thinking. Article explains the actual CHI survey's bounds.
limitations
Headline and comments are not causal evidence. No X access or professional consensus is claimed.

professional panel

name
Erik Brynjolfsson
credential
Stanford Jerry Yang and Akiko Yamazaki Professor; Stanford Digital Economy Lab director, as listed on Stanford Economics profile
official url
https://economics.stanford.edu/people/erik-brynjolfsson
public usefulness
Public digital-economy research, books and explanations help readers distinguish an aggregate productivity claim from a specific workplace deployment.
contribution
Coauthored the published support-agent study used for average and heterogeneous effects.
boundary
That study covers one firm's support assistant; his broader expertise does not establish every tool's effects.
reviewed at
2026-10-01
name
Danielle Li
credential
David Sarnoff Professor of Management of Technology, MIT Sloan, as listed by MIT
official url
https://mitsloan.mit.edu/faculty/directory/danielle-li
public usefulness
MIT's public augmentation guidance explains how to define success before adopting workplace AI.
contribution
Coauthored support-agent evidence and supplies the practical success-criteria framing.
boundary
Institutional implementation advice does not validate STUDlearn's exact example or promise savings.
reviewed at
2026-10-01
name
Lindsey Raymond
credential
Assistant Professor jointly in MIT Economics, EECS and the Schwarzman College of Computing, as listed by MIT
official url
https://economics.mit.edu/people/faculty/lindsey-raymond
public usefulness
Public research on AI, labor and human learning makes workplace effects understandable through actual data rather than demonstration alone.
contribution
Coauthored Generative AI at Work, including experience differences and evidence consistent with learning.
boundary
Profile is no longer described as an incoming appointment; the research still does not establish universal learning or employment effects.
reviewed at
2026-10-01
name
Fabrizio Dell'Acqua
credential
HBS AI Institute researcher with a Columbia Management PhD; official personal site lists an incoming Michigan Ross assistant professorship
official url
https://www.fabriziodellacqua.com/
public usefulness
Public research explanations and teaching describe why nearby-looking knowledge tasks can produce different AI effects.
contribution
Lead author of the consultant frontier experiment, checked in its 2026 journal version.
boundary
Incoming appointment is reported as listed, not asserted as already begun; task boundaries from the experiment are not today's universal capability map.
reviewed at
2026-10-01
name
Ethan Mollick
credential
Associate Professor of Management and co-director of the Generative AI Labs at Wharton, as listed by Knowledge at Wharton
official url
https://knowledge.wharton.upenn.edu/faculty/ethan-mollick/
public usefulness
Public explanations and Co-Intelligence help ordinary readers consider AI assistance and human judgment together.
contribution
Coauthored the consultant experiment supporting uneven task performance.
boundary
Public examples and personal tool experiments do not establish a prompt's reliability or replace independent checking.
reviewed at
2026-10-01
name
Hamsa Bastani
credential
Associate Professor of Operations, Information and Decisions at Wharton with a secondary Statistics and Data Science appointment, as listed by Wharton
official url
https://oid.wharton.upenn.edu/profile/hamsab/
public usefulness
Public research on human/AI decision systems and safeguarded tutoring clarifies assisted performance versus later learning.
contribution
Coauthored the published high-school mathematics trial comparing unrestricted and safeguarded GPT-4 tutors.
boundary
Classroom learning evidence is adjacent to adult workplace accountability; it does not prove workplace deskilling.
reviewed at
2026-10-01
name
David Autor
credential
Professor of Economics at MIT, verified through MIT faculty profile
official url
https://economics.mit.edu/people/faculty/david-h-autor
public usefulness
Accessible essays explain tasks, substitution, complementarity and demand as mechanisms of workplace change.
contribution
2015 automation review supplies the conceptual distinction between automating a task and eliminating an entire occupation.
boundary
Historical framework, not a current numerical employment forecast or reassurance against displacement.
reviewed at
2026-10-01
name
Paweł Gmyrek
credential
Senior Researcher in the ILO Research Department's Effective Labour Institutions work, as listed in ILO research repository
official url
https://researchrepository.ilo.org/esploro/profile/pawel_gmyrek/overview?institution=41ILO_INST
public usefulness
Public ILO reports explain occupational exposure in a way readers can separate from layoffs.
contribution
Lead author of the refined 2025 ILO–NASK exposure index.
boundary
Task ratings and exposure estimates concern potential change, not observed employer-specific outcomes.
reviewed at
2026-10-01
name
Janine Berg
credential
ILO Research Department Senior Economist, identified in the opened May 2025 report's author biography
official url
https://brasil.un.org/sites/default/files/2025-05/OIT-NASK-IAGen_WP140_web.pdf
public usefulness
ILO's public exposure reports connect technological change to working conditions and transition support.
contribution
Coauthored the refined exposure index and its transformation-versus-replacement interpretation.
boundary
Role is verified in the dated report biography, not independently confirmed as a current appointment; modeling does not establish net job loss.
reviewed at
2026-10-01
name
Elham Tabassi
credential
Director of the Artificial Intelligence and Emerging Technology Initiative and Senior Fellow at Brookings; former NIST chief AI advisor and AI RMF leader, as listed by Brookings
official url
https://www.brookings.edu/people/elham-tabassi/
public usefulness
Public risk-framework work explains responsibility, trustworthy systems and meaningful measurement without relying on vendor marketing.
contribution
Led the AI RMF that supplies the four-function and trustworthiness teaching.
boundary
Current Brookings role is distinguished from historical NIST leadership; voluntary framework is not law or certification.
reviewed at
2026-10-01
name
Arvind Narayanan
credential
Princeton Computer Science professor and Center for Information Technology Policy director, as listed on official faculty page
official url
https://www.cs.princeton.edu/~arvindn/
public usefulness
Public AI as Normal Technology explanations encourage readers to distinguish demonstrated capabilities from advertised possibilities.
contribution
Provides the public explanatory lens of evaluating what a system actually does; factual risk claims here rest on opened studies and NIST guidance.
boundary
Profile and public teaching are professional context, not validation of one employer's assistant or evidence of a universal model failure rate.
reviewed at
2026-10-01
name
Ifeoma Ajunwa
credential
Asa Griggs Candler Professor of Law at Emory and founding director of its AI and the Future of Work initiative, as listed by Emory
official url
https://law.emory.edu/faculty/faculty-profiles/ajunwa-profile.html
public usefulness
The Quantified Worker and public employment-technology scholarship help readers ask about privacy, monitoring and decisions affecting workers.
contribution
Keeps AI used to evaluate a worker distinct from AI chosen by a worker to draft a message.
boundary
Professional scholarship and profile are not jurisdiction-specific legal advice or a universal right-of-appeal claim.
reviewed at
2026-10-01
name
Stijn Broecke
credential
Senior Economist in OECD Employment, Labour and Social Affairs, leading future-of-work research, as listed by OECD.AI
official url
https://oecd.ai/en/community/stijn-broecke
public usefulness
Public OECD research explanations and talks connect workforce consultation and training to reported experiences.
contribution
Coauthored the 2023 AI workplace survey report used for training/consultation associations.
boundary
Early-2022 survey perceptions are not a randomized causal effect or a survey of current generative tools.
reviewed at
2026-10-01
name
Kate Kellogg
credential
David J. McGrath Jr. Professor of Management and Innovation at MIT Sloan, as listed by MIT
official url
https://mitsloan.mit.edu/faculty/directory/kate-kellogg
public usefulness
Public organizational research and MIT explainers describe how implementation, expertise and worker participation change technology's practical effects.
contribution
Coauthored consultant frontier research and contributes organizational context for training, work design and review.
boundary
Neither a profile nor implementation research guarantees a particular training package's causal effect.
reviewed at
2026-10-01
name
Reva Schwartz
credential
Named coauthor of NIST AI 600-1, the Generative Artificial Intelligence Profile (2024), verified on NIST's publication record
official url
https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence
public usefulness
Coauthored publicly accessible professional risk guidance explaining confabulation, privacy and human-AI configuration.
contribution
Contributes to the official profile supporting over-reliance and confidentiality risk teaching.
boundary
Professional authorship is verified; a current job was not independently confirmed. Coauthorship does not attribute every passage to this individual.
reviewed at
2026-10-01
name
Advait Sarkar
credential
Senior Researcher, Microsoft Research Cambridge, as listed on Microsoft Research profile
official url
https://www.microsoft.com/en-us/research/people/advait/
public usefulness
Public human-computer-interaction research helps readers understand verification and critical-thinking effort in AI-assisted work.
contribution
Coauthored the CHI survey of 319 workers and 936 examples used to clarify confidence associations.
boundary
Microsoft-employed researcher; self-reported associations are not causal skill-loss evidence. Profile warns it is not regularly updated, so role is stated as listed.
reviewed at
2026-10-01
name
Lev Tankelevitch
credential
Senior Researcher and behavioral scientist at Microsoft Research Cambridge; Oxford PhD background in cognitive psychology/neuroscience, verified on Microsoft profile
official url
https://www.microsoft.com/en-us/research/people/levt/
public usefulness
Public metacognition and human-agency explanations connect thinking about one's own judgment to AI-assisted collaboration.
contribution
Coauthored the critical-thinking survey and supplies relevant expertise for confidence versus verification distinctions.
boundary
Microsoft-employed researcher. Survey supports reported associations, not a universal deskilling effect or a validated workplace intervention.
reviewed at
2026-10-01
name
Shakked Noy
credential
MIT Economics PhD student, described on official personal site as fifth-year and on the 2026–27 job market
official url
https://shakkednoy.com/
public usefulness
Public coauthored VoxEU account explains writing-task design and the gap between an experimental result and workplace use.
contribution
Coauthored the 453-person Science writing experiment; the author's site links the manuscript used here.
boundary
The listed PhD is in progress. The short, generic writing tasks do not establish real-file accuracy or productivity across workplaces.
reviewed at
2026-10-01

Download original public profile Download complete article

End of AI at work