STUDboardsPlay · Research · Self · Fit · Money · Learn · Study · Plan · Token STUDtokenOptional

STUDlearn · Work

AI at work

Also known as

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 be wrong or incomplete. Not medical, legal, financial, or other professional advice.

Open in the directory ↗

Use a workplace AI tool for drafting and checking without handing it the job, the facts, or private material.

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

What would help today?

Start a work draft without handing over the job
A useful first task
Ask for a structure or a rewrite using public material.
Your part
Put the correct names, numbers, tone, and judgment back into the draft before sending.

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.

Sources 1
Check a fluent draft before anyone relies on it
Before anyone relies on it
Compare the draft’s facts and citations with the original email, current notice, spreadsheet, or other source you own.
How much checking?
Match the review to the cost of being wrong; fluent writing can still be incorrect.

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.

Sources 3
Keep confidential work out of tools you do not control
Use an approved route
Put restricted work into a tool only when workplace rules, professional duties, and authorization already cover that use.
Otherwise
Keep client identifiers, patient details, contracts, and secret figures out.

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. This page is not legal advice for one profession. 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.

Sources 4
Divide the job into tasks you still own
Divide one job
Name one task AI may help with, one check you still perform, and one part that requires your judgment, access, or sign-off.
What changes
Some tasks can change without the whole occupation disappearing.

A tool can take some drafting or lookup steps without owning the occupation. The International Labour Organization’s 2023 analysis of generative AI and jobs found exposure concentrated in some clerical and cognitive tasks, with the main expected effect closer to changing work than wiping out whole occupations, and with large uncertainty about net employment. David Autor’s longer automation research makes the same kind of distinction: machines often substitute for some tasks while raising the value of others. Pawel Gmyrek and Janine Berg warn against reading a global exposure score as a personal layoff notice. Name one task the tool may help with this week, one check you still perform, and one part of the job that still needs your judgment, access, or sign-off.

Sources 5

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.

What the research found

  • Treat AI as a task helper, not a whole-job replacement. Exposure means some of a job’s tasks could in principle be done or assisted by a system, not that the occupation disappears. Gmyrek, Berg and David Bescond found clerical work most exposed, with many other occupations only partly touched, and they interpret the central effect as likely augmentation rather than wholesale automation. Autor’s task view of technological change helps keep that picture: substitution in one step can coexist with remaining human work in others. Use the finding to sort this week’s tasks, not to forecast your employment. A global index is not a notice from your employer, a reason to skip verification, or permission to paste confidential material.
    Sources 4
  • 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.
    Sources 1
  • 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.
    Sources 5
  • 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.
    Sources 4
  • 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.
    Sources 4
  • 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.
    Sources 2

Where experts still disagree

  • 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.
    Sources 2
  • Keep human judgment visible when the task sits near the edge. Dell’Acqua’s consulting experiment found large quality and speed gains on tasks inside GPT-4’s then-current reach and worse correctness on a task selected to sit outside it. Brynjolfsson’s support-agent study found the opposite pattern on experience: newer agents gained more. Those designs answer different questions, so they do not crown one universal winner about who benefits. Ethan Mollick’s public translation of the jagged frontier is a practical caution: nearby-looking work is not equally automatable. When you cannot tell which side you are on, keep a human check that can still catch a fluent miss, and do not treat one experiment as a staffing rule.
    Sources 4
  • 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.
    Sources 2
  • Weigh augmentation against displacement without a personal forecast. ILO authors expect more task change than mass occupation loss, with clerical work and women’s employment in some high-income settings more exposed because of the mix of tasks. OECD respondents still reported knowing of job losses in some adopting firms and remaining worry about the next decade. Autor emphasizes complementarities that historically kept employment from collapsing even as tasks automated. These are different methods — occupational scoring, perception surveys, and historical labor-market research — so they can be read together without forcing a single percentage. Use them to plan skills, checks and job quality, not to announce that your role is safe or finished.
    Sources 3

Just talk

  • 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.
    Sources 2
  • 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.
    Sources 2
  • 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.
    Sources 3
  • 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.
    Sources 3

What to try

  • 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.
    Sources 3
  • 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.
    Sources 1
  • 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.
    Sources 2
    • 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.Research or guidanceDOI (opens in a new tab) ↗NIST publication record (opens in a new tab) ↗
    • 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.Research or guidanceSSRN (opens in a new tab) ↗
  • 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.
    Sources 3

How to 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.
    Sources 1
  • 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.
    Sources 3
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1

Sayings people repeat

  • 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.
    Sources 1
  • 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’s 2023 analysis found exposure concentrated in some clerical and cognitive tasks, with the main expected effect closer to augmentation than wholesale occupation loss and with large uncertainty about net employment. Autor’s automation research likewise separates task substitution from total job collapse. A forecast is not a personal layoff notice.
    Sources 2
  • 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.
    Sources 3
  • 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.
    Sources 2
    • 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.Research or guidanceDOI (opens in a new tab) ↗NIST publication record (opens in a new tab) ↗
    • 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.Research or guidanceSSRN (opens in a new tab) ↗
  • 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.
    Sources 2
  • 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.
    Sources 1
  • 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.
    Sources 3

The longer notes

  • 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.
    Sources 1
  • 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.
    Sources 3
  • 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.
    Sources 4
  • 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.
    Sources 4
  • Ask what a hiring or monitoring tool is actually deciding. Ajunwa’s research and public testimony describe automated hiring, video interview scoring and productivity monitoring as workplace technologies that can replicate or hide discrimination and extend surveillance. This page is not a compliance manual or a ruling in a particular case. The visitor job here is ordinary drafting, not building a hiring model. The transferable question is the same: who is accountable for the decision, and can a person challenge it? A generated ranking is not a lawful or fair conclusion by itself.
    Sources 1
  • 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.
    Sources 3
  • 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.
    Sources 4
  • Keep sister topics in their own jobs. Career and work covers burnout, psychological safety, job crafting and reversible career experiments, including a short ILO-based note on testing AI on a bounded task you already own. Learning and remembering covers retrieval, spacing and the Bastani classroom result as study support. Making sense of online information covers lateral reading and AI-image checks. This topic is the workplace use of generative tools for drafting, checking, privacy, over-trust and task division. Point across when the question changes. Do not collapse them into one chatbot lesson or one job-loss headline.
    Sources 2

Who this is drawing from

  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1
  • 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.
    Sources 1

Related

Good to know

  • 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.
    Sources 3
  • Not advice. Named sources. Honest paraphrase of the finding. Not medical, legal, or financial advice.

Also on this house

STUDlearn. For burnout, work design, and career experiments, see Career and work.

All STUDlearn topics →