A performance review has three parts: the evidence, the verdict and the words. The evidence and the verdict belong to the manager, every time. ChatGPT, Claude, Gemini and Copilot are useful for the words, and more useful still as a second reader that checks the words for the habits that make reviews unfair. Used that way, AI makes a review more specific. Asked for the verdict, or handed a description of a person instead of examples of their work, it makes the review more generic and often more biased.
What goes wrong when AI writes it
The usual request is a name, a role and an impression: “Write a review for Jordan, a marketing coordinator. She’s great but needs to be more confident.” The model has no examples, so it writes the most typical review for that role and that impression.
In 2023, Textio co-founder Kieran Snyder ran requests like this through ChatGPT and published what came back: feedback full of clichés, with no examples and little anyone could act on. The pronouns followed stereotypes. Nurses were presumed female nine times out of ten, and an employee described as confident was written about as a man three times as often as a woman. Those were informal tests on an early model, and current models are more careful. The mechanism has not changed: give a model an adjective about a person and it fills in the rest of the person.
- Nurses presumed female
- 9 in 10
- Employees described as confident, written about as men rather than women
- 3 to 1
Informal tests on an early model, published in 2023. Current models are more careful; the mechanism has not changed. Source: Kieran Snyder, Textio, 2023.
The same bias is in reviews people write
Before blaming the model, look at what it learned from. In 2014, Snyder collected 248 performance reviews from 180 people at 28 companies, all written before chat models existed. Critical feedback appeared in 58.9% of the reviews men received and 87.9% of those women received. Negative comments about personality appeared in 2 of the 83 critical reviews men received, and in 71 of the 94 critical reviews women received (Fortune, 2014).
Reviews written by managers, without AI, collected in 2014. A small, self-selected sample that was not peer reviewed, but the gap is large. Source: Kieran Snyder, Fortune, August 2014.
The sample was small and self-selected, and it is twelve years old. I cite it for one reason. It shows that the failure we worry about with AI, judging the person instead of the work, is a failure reviews already had. That turns the question into a practical one: which way of working pulls a review toward evidence?
Who does what
My answer is a split in five steps, with the person at both ends:
The model never sees an impression of the person, only examples of their work, and never chooses the rating.
Two details carry most of the weight. The rating is decided before the model sees anything, so fluent prose cannot talk you into a different verdict. And the model gets examples rather than adjectives, so it has nothing to fill in.
The request, before and after
When we wrote the performance review template for our prompt repository, we built it on those two rules: it asks for the rating you intend to give, leaves every judgment about rating and pay to you, and ties each claim to an example from your notes. Here is a shorter version of the same idea next to the usual request, with scores from the free checks in the Prompt Grader:
Before
2/10 Needs work
- Has: What to make
- Missing: Who it's for
- Missing: Background facts
- Missing: Length
- Missing: Format
- Missing: What to avoid
- Has: What good looks like
- Missing: An example
- Missing: If unsure
- Missing: Layout
After, with what was added marked
9/10 Strong
- Has: What to make
- Has: Who it's for
- Has: Background facts
- Has: Length
- Has: Format
- Has: What to avoid
- Has: What good looks like
- Missing: An example
- Has: If unsure
- Has: Layout
The second version states the rating already decided, works only from dated examples, rules out comments on personality and marks every gap with [no example]. The marked words are what it adds.
The “[no example]” line matters most. It turns a gap in your notes into something you can see, instead of a confident sentence the model wrote to cover it.
AI as the second reader
The job I would give AI even if you write every word yourself is the audit. Paste your draft and ask for three lists: sentences that would fit anyone in the same role, sentences about personality or attitude rather than actions, and claims with no example behind them. Then fix them yourself. A typical catch looks like this:
Jordan needs to be more confident and assertive in meetings.
In the March and June campaign reviews, Jordan had the numbers that settled the budget question but shared them only when asked. Next quarter: open each review with your numbers.
The second version is longer and kinder, because it describes something Jordan can repeat or change. The specific feedback template does this rewrite for any piece of feedback, and the Robot-Speak Checker catches the stock phrases and filler a model adds on top.
1/10 Needs work
- Has: What to make
- Missing: Who it's for
- Missing: Background facts
- Missing: Length
- Missing: Format
- Missing: What to avoid
- Missing: What good looks like
- Missing: An example
- Missing: If unsure
- Missing: Layout
Ask for the three lists: sentences that would fit anyone in the same role, sentences about personality rather than actions, and claims with no example behind them. Say who reads the review, and that you will make the fixes yourself. Open in the Prompt Grader
What stays out of the chat
A review is personnel data. Use the AI account your company has approved for it, not a personal one, and take names and identifying details out before you paste. Safe Paste swaps them for placeholders in your browser and puts them back in the answer. Leave out medical, family and legal matters entirely; they do not belong in a review draft anyway. The AI at work risk checklist covers the rest.
The best objection
The strongest objection is that a review is the one document an employee should receive in their manager’s own words, and that using AI at all says the review was not worth the manager’s time. I take that seriously. It is why I would never let a model choose the rating or supply an example.
But the words were never the hard part of a review. The hard part is having kept the evidence and being specific about it. A manager writing alone from memory on a late November evening is more likely, not less, to produce the generic, personality-flavored review the 2014 numbers describe. The test I would apply is simple: if your employee asks how the review was written, the honest answer should be one you are comfortable giving.
The same split works for the other side of the table. The self-assessment template drafts from the employee’s own examples, and the review season pack collects eleven prompts for the season, from 1:1 prep to pay conversations. If your team shares one review prompt, give it the ten-minute prompt review before the season starts.
Paste your impression of the person and ask for a balanced review with a suggested rating
Not quite. Given an impression, a model writes the most typical review for it and fills in the rest of the person. The rating is the manager’s call, made before any AI sees the notes.
Decide the rating first, then give the model your dated examples with names removed, and ask it to cite one for every point
Yes. The verdict is settled before fluent prose can argue with it, and examples leave the model nothing to fill in. Where an example is missing, it writes [no example].
Leave AI out, since reviews people write themselves are free of this bias
Not quite. The 2014 reviews were written before chat models, and 71 of 94 critical reviews women received still criticized personality. An AI audit pass is one way to catch those lines.
The checklist
- Dated examples for every point, kept through the year rather than reconstructed in November.
- The rating decided before any AI sees the notes.
- An approved work account, with names and identifying details replaced.
- A draft that cites an example for every strength and every area to work on.
- An audit pass that flagged sentences fitting anyone, and sentences about personality.
- Every sentence one you would defend in the meeting.
Key takeaways
- The evidence and the rating belong to the manager; AI is useful for the words and for checking them.
- Given an impression instead of examples, a model writes the most typical review, and informal tests found its pronouns following stereotypes.
- Reviews written without AI had the same flaw: in a 2014 sample, 71 of 94 critical reviews women received criticized personality, against 2 of 83 for men.
- Decide the rating first, give the model dated examples only, and let it write [no example] where your notes have a gap.
- Use AI as a second reader that lists generic lines, personality comments and claims without evidence.
- Use an approved work account and take names out before you paste.
Questions
Is it OK to use ChatGPT to write a performance review?
Yes, for drafting and checking, if your company allows it and the judgment stays yours. Decide the rating first, give the model dated examples rather than impressions, use an approved work account and take names out. Then edit until every sentence is one you would defend in the review meeting.
What should I not put into AI when writing a performance review?
Names and identifying details, and anything medical, family-related or legal. Use the AI account your company approves for personnel data, and replace names with placeholders first. Safe Paste does that in your browser and puts the names back in the answer.
Can AI make performance reviews less biased?
It helps most as a checker. Ask it to list sentences that would fit anyone in the role, sentences about personality rather than actions, and claims with no example behind them. Used as the writer from a short description of a person, it tends to repeat stereotypes instead, as informal 2023 tests of ChatGPT by Textio found.
How do I write my self-assessment with AI?
The same way: collect dated examples of what you did and what came of it, decide what you want your manager to take away, then ask the model to draft from those examples only. The self-assessment template is set up to work that way.
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.
- Respond in writing to a performance review you disagree withYour performance review has a rating or comments you think are wrong, and you want a calm written response based on facts to submit or bring to a meeting.
- Write your self-assessment for a performance reviewYou need to summarize your year with evidence, honest gaps and goals before review season.