A 360 review request lands in your inbox asking for feedback on a colleague, or your manager asks you to write up specific feedback for someone you supervise. You have a general sense of how it’s gone - mostly good, with one recurring friction point - but turning that general sense into clear, specific, useful written feedback is harder than it sounds. AI is a real help here for structure and phrasing. It is not a source of the actual content, and treating it as one is where this goes wrong.
This is a different problem from refusing to rank people with AI, which covers the specific danger of asking a model to score or rank a person. This article assumes you are not asking for a score - you are writing qualitative feedback about specific behavior, and the risk is different: AI smoothing your actual observations into something vaguer, harsher, or more generic than what you actually witnessed and intended to say.
The rule: examples are yours, structure is AI’s
Every specific claim in a piece of feedback - “missed the deadline on the March report,” “pushed back constructively in the planning meeting,” “interrupted colleagues repeatedly in stand-up” - has to be something you actually observed, ideally with a rough date or instance attached. AI has no independent access to what happened; if you ask it to “write feedback about someone who could communicate better,” it will produce plausible-sounding, generic examples that read as specific but are not tied to anything real. That is fabrication dressed as detail, and it is exactly what makes feedback feel unfair to the person receiving it - a vague or invented example cannot be discussed, corrected, or even properly understood, because there is nothing concrete behind it.
Never let AI generate an example, incident, or specific behavior in a piece of feedback about a real person. If you cannot recall a concrete instance to support a point, either find one before writing the feedback or drop the point - a general impression without a specific example is not useful feedback, and an invented one is worse than none.
A workflow that keeps examples real
Step 1: List your own specific observations first
Before opening any AI tool, write down what you actually remember: specific situations, what the person did, and what the effect was. The Center for Creative Leadership’s widely used Situation-Behavior-Impact model is a useful shape for this even before AI enters the picture - situation (when/where), behavior (what specifically happened, described neutrally), impact (what effect it had) (Center for Creative Leadership, “Use SBI (Situation-Behavior-Impact) to Inquire About Intent”).
Step 2: Use AI to organize and phrase, not to generate content
Here are specific observations I want to give as feedback to a
colleague: [paste your situation-behavior-impact notes]. Help me
phrase these clearly and professionally using this same structure.
Do not add any example, behavior, or impact that I did not
describe above. Keep the tone matched to what I actually intend -
do not make constructive points sound harsher than I described them,
and do not soften a genuine concern into vague praise.
The second instruction matters as much as the first - a model asked to “make this diplomatic” will sometimes round a real, specific concern into a euphemism vague enough that the person receiving it cannot tell what to actually change.
Step 3: Check the tone against your actual intent, in both directions
Read the draft back and ask two separate questions: does this sound harsher than what I actually meant to say, and does this sound softer or vaguer than what I actually meant to say? Both directions cause real damage - overcorrected harshness can misrepresent a minor issue as a serious one, and overcorrected softness can bury a real, actionable concern so deep in diplomatic language that the person never registers it as something to change.
Step 4: Verify every example against your own memory before sending
For each specific claim in the draft, confirm you can actually recall the instance it refers to. If a sentence reads well but you cannot place the specific event behind it, that is a sign the drafting process let something generic slip in - cut it or replace it with something you can actually stand behind.
If the feedback will feed into a decision about pay, promotion, or a formal performance process, use your organization’s official feedback form or channel rather than an informal AI-drafted note, and follow whatever process your HR team has defined for that specific context - a well-phrased private message is not a substitute for the record a formal process requires.
What AI is and isn’t good for here
| Task | Safe | Risky |
|---|---|---|
| Organizing your own observations into a clear structure | Yes | - |
| Phrasing a point more professionally without changing its substance | Yes | Watch for softened-into-vague drift |
| Checking tone consistency across multiple points | Yes | - |
| Generating an example or incident you did not supply | No | This is fabrication |
| Deciding how harsh or soft the feedback “should” be | No | That judgment is yours, based on the actual behavior and its actual impact |
| Describing a person’s characteristics rather than specific behavior | No | Focus on behavior, not personality labels - a described action can be discussed and changed, while a characteristic often cannot, and staying on behavior keeps you away from language that reads as commentary on something protected. The U.S. EEOC’s resources on AI and the ADA address automated employment decision tools rather than informal drafting, but the underlying caution transfers |
Why this differs from a performance review or a ranking
It is worth being precise about what this article does and does not cover, because the three are easy to blur. Writing your own performance review self-assessment is about organizing evidence of your own work. Refusing to rank people with AI is about never letting a model produce a score, grade, or “who is better” verdict about a person. This article sits between them: you are producing qualitative, evidence-based observations about someone else’s specific behavior, for their benefit, not a number and not your own record. The accountability rule is the same across all three - a named human owns the content - but the material and purpose differ enough that treating them as one workflow leads to mistakes in each.
Polish is not fairness
It is easy to assume that AI-polished feedback is automatically fairer feedback, because it reads calmer and more structured than a first draft written in frustration or haste. Polish and fairness are different things. Feedback is fair when it is specific, evidence-based, and matches what actually happened - not when it merely sounds professional. A calmly worded piece of feedback built on a vague or invented example is not an improvement on a rougher one built on something real; it is a better-dressed version of the same problem.
Draft your next piece of feedback
Before your next 360 review or feedback conversation, use the workplace feedback draft checklist to capture your real observations first, then draft with AI for structure and tone only. If the feedback is part of preparing for a 1:1 conversation, 1:1 agenda prep covers the same in-your-own-words discipline for how you actually deliver it.
