What AI should and should not do in a review
A review has two parts that feel like one: the evidence of what someone did over the year, and your judgment of what it adds up to. AI can help a great deal with how the evidence is arranged and worded. It cannot supply the evidence, and it should not make the judgment.
| Let AI do it | Keep it yourself |
|---|---|
| Sort a year of rough notes into the sections of your form. | Decide which notes matter and what they show. |
| Group examples into themes, such as “client work” or “process fixes”. | The rating, and any link to pay or promotion. |
| Turn a blunt note into clear, specific, kind wording. | Anything about health, family, leave or personal life (leave it out). |
| Suggest goals that follow from the evidence. | Comparisons with teammates. |
| Point out claims that have no example behind them. | The final wording the person reads. |
Keep that split and the review stays yours. Blur it, by asking the model to “write a review for a solid performer”, and you get a document that sounds finished and says nothing the person can act on.
- 1Collect evidence firstDated notes from the whole year, not the last month.
- 2Draft from your notes onlyEvery judgment tied to an example.
- 3Check for vague and loaded wordsReplace each with what the person did.
- 4Make it yours, and set the ratingRewrite in your voice. The rating is your call.
1. Collect evidence before you open the chat
The quality of an AI-drafted review is set before you type the prompt. Most managers sit down in December with a clear memory of October and a fuzzy one of February, and the review leans toward whatever happened recently. A model cannot correct that. It only sees what you give it.
The fix is unglamorous: a line or two of notes after each one-on-one, with the month, what the person did and what came of it. “Mar: rebuilt event sign-up form, registrations up from 140 to 210” is worth more than a page of adjectives in December. The one-on-one agenda template leaves room for exactly this, and the specific feedback template turns a note into feedback you can give in the moment, so the review holds no surprises.
If you did not keep notes, gather what exists: project updates, shipped work, customer emails, peer comments. Then be honest in the prompt about which months are thin. A model told “I have little from January to April” writes a review that says less about those months, which is better than one that invents.
2. Draft from your notes only
The prompt has one job: stop the model from adding anything you did not give it. That means telling it to use only your notes, to tie each judgment to a dated example, to leave the rating out, and to say when a section has no evidence instead of filling it with pleasant generalities.
Write a performance review for my employee, a marketing coordinator. They did a good job this year.
I manage a marketing coordinator and need a first draft of their annual review. It goes into our HR system, where they and my director will read it.
My notes from the year, with dates:
"""
{{notes}}
"""
Use the three headings from our review form: Results, How they work, Goals for next year.
Write each section in 80 to 120 words.
Use only my notes. Every judgment must point to a specific example from the notes, with its month.
Example of the detail I want: "In March, they rebuilt the event sign-up form, and registrations rose from 140 to 210."
Do not suggest a rating; I will decide it.
Do not comment on personality, age, family or health, and avoid labels such as "abrasive" or "emotional" unless you describe the specific behavior behind them.
A good draft is one they could read and know exactly what to keep doing and what to change.
If my notes do not cover a section, write [NO EVIDENCE] instead of guessing.What changed. “They did a good job” gives the model nothing, so it writes the most typical review for a marketing coordinator: collaborative, detail-oriented, a team player. Every word of it could describe anyone. The second version gives it only evidence, and makes the gaps visible as [NO EVIDENCE], which tells you where your notes are thin before your employee notices.
Scores from the Prompt Grader’s free ten checks, run on October 28, 2026.
The performance review template is a fuller version of this brief, with blanks for your review guidelines, goals, self-review and peer feedback, and a “Check before sending” list after the draft. If the review leads to a formal improvement plan, use the improvement plan template rather than stretching a review prompt to cover it; a plan needs measurable goals and dates, not a narrative.
3. Check the draft for vague and loaded words
Vague feedback is unhelpful, and it is also where bias hides. In a Stanford study of reviews at four companies, reported in Harvard Business Review, women were less likely than men to receive specific feedback tied to business outcomes, whether the feedback was praise or criticism. Without specifics, a strong year is hard to turn into a promotion case.
Language models carry patterns like this from the text they learned from. Researchers who asked models to write reference letters found significant gender bias in both the style and the words used, and warned that such letters could hurt women’s applications. Nobody has published a study of AI-written performance reviews specifically that we could find, so treat this as a warning sign rather than a measured effect. The practical lesson is the same either way: the more the model writes from its own sense of a “typical” employee, the more of these patterns get in.
Read the draft once with only this in mind, and mark three kinds of phrase:
- Praise with no example: “strong communicator”, “great attitude”, “a pleasure to work with”. Ask for the example, or cut it.
- Labels about personality: “abrasive”, “emotional”, “not confident”, “too quiet”. Replace each with the specific behavior and its effect, or drop it.
- Stock review language: “exceeds expectations in all areas”, “a valued member of the team”. These read as AI-written, and they are. The Robot-Speak Checker flags the stock phrases for you.
A follow-up prompt that does the first pass: “List every sentence in the draft that has no dated example from my notes behind it, and every word that describes personality rather than behavior.” Then fix them yourself.
4. Make it yours, and set the rating
The person reading the review will read it closely, probably more than once. It should sound like you, the manager who sat in their one-on-ones, not like a template. Rewrite the opening and closing of each section in your own words and keep the AI’s structure where it helps. Research on workplace email has found that people judge senders who lean heavily on AI as less sincere, and a review is the most personal document most managers write all year. The guide to writing emails with AI covers that research.
The rating stays with you. It drives pay, promotion and sometimes whether someone keeps their job, and a model has neither the full picture nor any accountability for it. That is also why the prompt says “Do not suggest a rating”: once a number appears on the screen, it is hard not to anchor on it.
What not to paste
A review draft is personnel data. Before you paste notes into any AI tool, check your company’s AI policy and use the workplace account it approves, if there is one; a personal ChatGPT, Claude or Gemini account is the wrong place for it. Either way, the model does not need the person’s name, and it should never see health details, family circumstances, leave, or anything from a disciplinary or legal process. Replace the name with “they” or a role, and run the notes through Safe Paste to catch the rest. The risk checklist for AI at work puts reviews among the jobs that always need a person to sign off.
Writing your own self-assessment
The same rules apply in reverse, and the temptation is the opposite one: AI makes everyone sound like a top performer. Start from your own list of what you shipped, fixed and learned, with dates and numbers, and ask the model to organize it under your form’s questions without adding anything. The self-assessment template does this, and asks what did not go well and what you changed as a result, which reviewers trust more than a page of strengths. If the review is also your moment to ask for more, the raise or promotion template builds the case from the same evidence.
A checklist to copy
Before you prompt [ ] Notes cover the whole year, with months [ ] Thin months named honestly [ ] Name, health, family and leave details removed [ ] Using the AI tool your company approves In the prompt [ ] "Use only my notes" [ ] Every judgment tied to a dated example [ ] Your form's section headings and lengths [ ] "Do not suggest a rating" [ ] "Write [NO EVIDENCE] instead of guessing" Before it goes in the system [ ] Every claim has an example behind it [ ] No personality labels, only behavior and effect [ ] Stock phrases cut [ ] Opening and closing rewritten in your voice [ ] Rating decided by you, without the draft open
Key takeaways
- Let AI organize and word your evidence. Keep the evidence and the judgment yourself.
- Notes kept through the year are what make an AI draft fair; the model only sees what you give it.
- Tell it to use only your notes, tie each judgment to a dated example and leave the rating out.
- Vague praise and personality labels are where bias hides. Replace them with behavior and results.
- Remove names and sensitive details, use an approved tool, and rewrite the draft in your own voice.
Questions
Is it OK to use ChatGPT to write a performance review?
Yes, as a drafting aid, if your company’s AI policy allows it and the evidence and rating come from you. Give the model your dated notes, tell it to use only those, and rewrite the draft in your own words. What is not OK is asking it to write a review from a job title and a feeling, because it will fill the gaps with generic praise that could describe anyone.
Can employees tell if a review was written by AI?
Often, yes. AI-written reviews tend to use the same stock phrases, “a valued member of the team”, “consistently exceeds expectations”, and few dated examples. Research on workplace email found that people rate senders who rely heavily on AI as less sincere. A review built from your own notes and rewritten in your voice does not read that way.
Can I paste employee information into ChatGPT?
Check your company’s AI policy first, and use the workplace account it approves if there is one. Even then, leave out the person’s name, health, family, leave and anything from a disciplinary or legal process. The model needs the work, not the identity. Safe Paste swaps names and other details for placeholders before you paste.
How do I use AI to write my self-evaluation?
Start from your own dated list of results, then ask the model to organize it under your form’s questions without adding anything. Ask it to point out claims that have no number or example behind them, and include what did not go well and what you learned. The self-assessment template is set up this way.
Should AI suggest the performance rating?
No. A rating affects pay and promotion, the model sees only part of the picture, and any number it suggests will anchor your own judgment. Decide the rating from the evidence yourself, ideally before you read the AI draft, and tell the model not to suggest one.
Prompts to use with this
- Draft a performance review from your own notesReview season is here and you have a year of notes, a self-review and peer feedback, and you need a fair, specific first draft to edit.
- Write your self-assessment for a performance reviewYou need to summarize your year with evidence, honest gaps and goals before review season.