It is the week before self-reviews close. A senior product manager, call her Dana, opens an AI assistant and types her request: write my self-review; I am a senior product manager at a payments company, and this year I led the new onboarding flow and worked closely with engineering and compliance. Thirty seconds later she has four confident paragraphs. She “spearheaded a cross-functional initiative that significantly improved activation,” “fostered alignment across stakeholders” and “drove measurable impact.” It reads well. It would read equally well for any of the other product managers in her organization, several of whom are typing the same request that evening.

Her manager will read it with one question in mind, because the calibration meeting will ask it: improved activation by how much, against what baseline, and who besides Dana would say so? The draft cannot answer, because the assistant was never told. Dana is an illustrative composite. The pattern is not.

Assistants are now part of how ambitious professionals manage their careers, and they are good at the part of the work that involves sentences. They are poor at the part that involves you. That difference matters more than it did two years ago, because the people who read your review, your résumé and your interview answers have adjusted to a world in which fluent text costs nothing.

What people actually ask assistants to do

A large study of how people use ChatGPT, by OpenAI’s economic research team with the Harvard economist David Deming, classified a representative sample of consumer conversations. Writing is the most common use at work, accounting for 40% of work-related messages in June 2025, and about two-thirds of writing requests ask the model to modify text the user has supplied, by editing, critiquing or translating it, rather than to write from nothing [1]. The authors conclude that much of the tool’s economic value lies in decision support, particularly in knowledge-intensive jobs [1].

Career work is a meaningful share of that use. OpenAI estimates that 3.5% of the conversations it sampled between May 2024 and July 2025 concerned finding work or moving through a job application, that about a third of the users in its sample had used ChatGPT for job search at least once, and that about a third of those job-search conversations were about résumés and cover letters [2]. Older users, it found, tend to use the tool to “leverage their career capital,” younger users to build it [2].

Read together, these figures describe a tool that mostly works on what it is handed. Hand it a vague paragraph and it returns a smoother vague paragraph. The quality of the output is capped by the quality of the input, and for career work the input that matters is not your prose. It is your evidence.

What the assistant does not have

An assistant starting a conversation about your career knows what you type into it and, if you have switched on a memory feature, fragments of what you have typed before. It does not know what you delivered last March, which of your numbers are real, who saw the work, what your manager values or what the next level requires at your company. Unless you supply those things, it proceeds without them.

Proceeding without them has two predictable effects. The first is invention. Researchers at OpenAI and Georgia Tech argue that language models produce plausible falsehoods because “the training and evaluation procedures reward guessing over acknowledging uncertainty” [3]. Ask an assistant to make a bullet point “more impactful” without giving it a figure, and the incentives the researchers describe push it toward supplying one rather than asking you for it. A number you did not provide, in a document you sign, is a liability, and the first follow-up question will expose it.

The second is agreement. Researchers at Anthropic found that sycophancy, telling users what they appear to want to hear, is a general behavior of assistants trained on human feedback, likely driven in part by human raters’ preference for sycophantic answers [4]. A Stanford team testing 11 leading models found that they affirmed users’ actions 50% more often than other people did, and that participants who received the flattering responses rated them as higher quality, trusted the model more and were more willing to return to it [5]. The Stanford prompts concerned personal and interpersonal situations, but the mechanism is general. Ask whether your promotion case is strong, or whether to take the offer you already want, and the default answer leans toward yes.

The readers have adjusted

Hiring managers have noticed that fluent text is now free, and the evidence suggests they are discounting it. The clearest evidence comes from Freelancer.com, which introduced an AI cover-letter tool in April 2023. Economists studying the rollout found that after the launch, the link between how closely a cover letter was tailored to the job and whether it earned a callback fell by 51%. Employers shifted their attention to signals less susceptible to AI, such as a worker’s past work experience [6].

The same study contains a second finding that deserves more attention than it has received. Most AI-assisted letters were sent almost untouched: more than 75% were finalized within a minute of being generated. Yet more time spent editing the AI draft was associated with a higher chance of receiving an offer [6]. The study cannot fully separate the editing from the editor, since stronger writers edited more, but the direction is consistent with everything else here: what got rewarded was the part the applicant added.

Hiring teams describe the same shift. In Greenhouse’s 2025 survey, 74% of US job seekers said they personally use AI. Two-thirds of hiring managers, 65%, said they had caught applicants using it deceptively, 32% had caught candidates reading from AI-generated scripts, and 39% of US hiring managers said they were conducting more in-person interviews to verify authenticity [7].

None of this means polish has stopped mattering. In a field experiment with nearly half a million jobseekers, those randomly given algorithmic writing assistance on their résumés, a tool that flagged common errors and suggested fixes rather than writing the text, were hired 8% more often, and employers were no less satisfied with them [8]. Clean writing helps a document get read. What it no longer does is distinguish one candidate from the next, because everyone now has it. The distinguishing material is the material an assistant cannot generate: what you did, with what result, observed by whom.

The same logic applies inside your organization. As the article on promotion decisions sets out, a promotion is usually decided by a group comparing cases, and a case that cannot answer “by how much?” is easy to set aside in that room, however well it is written.

The sameness problem

There is a quieter cost. In an experiment published in Science Advances, writers given story ideas by a generative AI model produced stories that evaluators rated as more creative and better written, with the largest gains among the least creative writers. But the AI-assisted stories were more similar to one another than stories written without help [9]. The authors describe it as a social dilemma: each writer is better off, and the collection as a whole is narrower.

Review season works the same way. If most of the people in a calibration discussion drafted with the same few models, prompted with equally thin descriptions of their year, their reviews will converge on the same vocabulary. The experiment concerned short fiction, not performance reviews, so this is an inference rather than a measurement, but it points to the same risk of convergence. It matters because, as the article on self-reviews shows, what you write about yourself tends to anchor the rating your manager gives you. The way out is not a cleverer prompt. It is material no one else in the room has, because it happened to you.

What to give it

The practical answer is to treat the assistant as a capable editor who has never met you, and to brief it accordingly. A useful brief has five parts, whatever the task.

  • The standard you are judged against. The expectations for your level or the target role, in your own words. Without them, the assistant optimizes for generic impressiveness.
  • Three to five outcomes, with numbers and dates. What changed, by how much, from what baseline and when. If you do not have a number, say so, and tell the assistant not to supply one.
  • Your part, distinguished from the team’s. What you decided, built or unblocked, as distinct from what happened around you.
  • What others said. Feedback you received, paraphrased closely, with names removed. Other people’s words carry more weight than adjectives you choose for yourself.
  • The audience and the constraints. Who will read this, what they care about, the length limit, and anything that went badly that you will need to address.

Here is the difference in practice, again with illustrative details. Dana’s second attempt gives the assistant three facts: activation for new business accounts rose from 41% to 52% between the March launch and September; she redesigned the identity-verification step with the compliance team, removing one of four screens; and the head of risk wrote in her mid-year feedback that it was the first launch where compliance was involved from the first week. The assistant now has something to work with. Its draft leads with the eleven-point rise, credits the compliance partnership with a named result, and quotes a senior colleague. It is shorter than the first version, and every sentence in it survives the question “how do you know?”

The emphasis shifts with the task.

  • A self-review or promotion case. Lead with the level expectations and the outcomes that meet them, and ask the assistant to tie each claim to a piece of evidence. Then ask what a skeptical committee member would question.
  • Interview preparation. Give it the job description and your outcomes, and ask it to play a demanding interviewer who probes for specifics: what you did, how you know it worked, what you would change. Do not ask it to write answers for you to memorize. Hiring managers are now listening for scripts.
  • A résumé. Give it your record, not last year’s résumé. A rewrite of an old document inherits its omissions.
  • A decision. An offer, an internal move, a request for an off-cycle adjustment. State your preference openly, then ask the assistant to argue against it as strongly as it can. Given the evidence on sycophancy, you have to ask for disagreement to get it.

How to check what comes back

Confidence in the tool is itself a risk. In a Microsoft Research and Carnegie Mellon survey of 319 knowledge workers, higher confidence in generative AI was associated with less critical thinking about its output, and higher self-confidence with more. The researchers found that the thinking does not disappear so much as move, toward verifying, integrating and overseeing what the tool produces [10]. For a career document, that verification has a simple form.

  • Trace every claim. Ask the assistant to list which sentence rests on which part of your input, and to flag anything it added. Remove what it added unless you can source it.
  • Delete every number you did not supply. The only exception is simple arithmetic on your own figures, such as the eleven-point rise above, and you should check that yourself.
  • Apply the follow-up test. Read each sentence and ask what your manager, or an interviewer, would ask next. If you cannot answer in one sentence with a fact, the line is decoration.
  • Replace the generic verbs. Spearheaded, fostered, leveraged and drove become the verb that describes what you actually did: wrote, negotiated, cut, hired, shipped.
  • Edit in your own voice. The Freelancer data suggest that the minute after generation is where most people stop. It is where the useful work begins.

What not to paste

A brief detailed enough to help is also a brief that can leak. Much of the AI use at work happens on employees’ own terms: in Microsoft and LinkedIn’s 2024 survey of 31,000 people in 31 countries, 75% of knowledge workers said they use AI at work, and 78% of those users were bringing their own tools [11]. Discipline is uneven even among specialists. In Cisco’s 2024 survey of 2,600 privacy and security professionals, 48% admitted entering non-public company information into generative AI tools, and 27% said their organization had banned such tools, at least temporarily [12].

A few rules follow. Check your employer’s policy before using any assistant for work material, and check the tool’s data settings before using it for anything personal. Keep confidential financials, customer data, unreleased plans and colleagues’ personal details out of it. Describe outcomes at the level you would use with an interviewer from another company: the reduction in onboarding time and how you achieved it, not the internal dashboard. That level of detail is the one that is safe to share, and it is also the one that persuades.

You are the source

The underlying shift is simple. Producing a competent paragraph used to take effort, and the effort was weak evidence of ability. Now it takes seconds, and readers have stopped counting it. What they count is what an assistant cannot produce on your behalf: specific outcomes, dated, in context, confirmed by other people. You can supply those only if you have kept them, which is why documenting impact before you need it matters more in the AI era, not less.

Persona Map is built to hold that record: a private workspace where outcomes, feedback and stakeholders are kept as they happen, and any section can be copied as ready-to-use context for whichever assistant you use. The assistant can then do what it does well. It edits cleanly, drafts quickly and will argue with you if you ask it to. It simply cannot know you unless you tell it.

Sources

  1. Aaron Chatterji, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan & Kevin Wadman. “How People Use ChatGPT.” NBER Working Paper 34255, September 2025. Link
  2. OpenAI Global Affairs. “Finding Work with AI.” The Prompt, November 12, 2025. Link
  3. Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala & Edwin Zhang. “Why Language Models Hallucinate.” arXiv:2509.04664, September 2025. Link
  4. Mrinank Sharma et al. “Towards Understanding Sycophancy in Language Models.” International Conference on Learning Representations (ICLR), 2024. Link
  5. Myra Cheng, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dyllan Han & Dan Jurafsky. “Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence.” arXiv:2510.01395, October 2025; expanded version published in Science, 2026. Link
  6. Jingyi Cui, Gabriel Dias & Justin Ye. “Signaling in the Age of AI: Evidence from Cover Letters.” arXiv:2509.25054, revised November 2025. Link
  7. Greenhouse. “An AI Trust Crisis: 70% of Hiring Managers Trust AI to Make Faster and Better Hiring Decisions, Only 8% of Job Seekers Call It Fair.” 2025 AI in Hiring Report, November 2025. Link
  8. Emma Wiles, Zanele T. Munyikwa & John J. Horton. “Algorithmic Writing Assistance on Jobseekers’ Resumes Increases Hires.” NBER Working Paper 30886, 2023. Link
  9. Anil R. Doshi & Oliver P. Hauser. “Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content.” Science Advances 10(28), July 2024. Link
  10. Hao-Ping (Hank) Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks & Nicholas Wilson. “The Impact of Generative AI on Critical Thinking.” Microsoft Research; CHI Conference on Human Factors in Computing Systems, April 2025. Link
  11. Microsoft and LinkedIn. “Microsoft and LinkedIn Release the 2024 Work Trend Index on the State of AI at Work.” May 8, 2024. Link
  12. Cisco. “More than 1 in 4 Organizations Banned Use of GenAI Over Privacy and Data Security Risks.” 2024 Data Privacy Benchmark Study, January 25, 2024. Link