The self-appraisals landed in the shared folder on Friday afternoon, and you opened the first one with real goodwill. It said the writer had "driven cross-functional alignment across a complex stakeholder landscape and delivered measurable impact on key deliverables." The second one said nearly the same thing in a different order. By the fifth you noticed that every one of them had three tidy headings, a confident closing paragraph, and not a single sentence that sounded like the person you work with.
Ratings are due in a week. You have eleven of these to read, and you are quietly wondering whether you are about to give a high score to a chatbot.
Here is my view, and I hold it firmly. Do not try to work out who wrote what. Detection is a guessing game that you will lose in one direction or the other, and either mistake lands on a real person. Somebody who used a tool to polish honest work gets treated like a cheat. Somebody who typed every word themselves, in careful corporate English, gets accused of sounding like a robot. Neither outcome is fair, and neither is yours to cause over a paragraph of style.
What I'd do instead is simpler. Read each appraisal for evidence, and ask for the evidence when it is missing.
Why polish tells you nothing this year
For a long time, a well-written self-appraisal was a loose signal of effort and communication skill. That signal is gone. A smooth paragraph now costs about thirty seconds, and it costs the same for the colleague who had a great year and the one who did not.
So the prose has stopped being information. What still carries information is anything the tool could not have known: a particular decision, a particular figure, a particular regret. A language model can write "improved turnaround time" about anyone. It cannot know that you pushed the Henderson order out a week to protect the audit, or that the backlog dropped from the number you remember to the number you hoped for. Only the writer knows those, and a person who did the work finds them easy to supply.
That gives you a clean test that does not depend on guessing anything about how the document was produced. Ask for the part only the person could know, and see whether it comes back quickly and specifically.
The one move: three questions, sent before the 1:1
Send each person the same short message, on Monday, a few days ahead of the conversation. Make it routine, so nobody feels singled out. Here is the one I'd use, ready to copy:
Thanks for sending your self-appraisal. Before we meet, I'd like three specifics to go with it. Short answers are fine, and a few lines each is plenty.
1. The one decision from this year you'd point to, and what the alternative was.
2. The one number you own, where it started, where it is now, and what moved it.
3. One thing you would do differently if you got the year back.
I'll read your appraisal alongside these when I write my rating.
It takes five minutes to send to the whole team, and it quietly changes what the appraisal is. The long document becomes the summary, and the three answers become the exhibits.
What each question is really checking
The decision. The question I'd ask myself while reading is whether there was a real fork. A strong answer says "we could have shipped on the date or held it for the fix, and I held it, and here is what that cost us." A thin answer names a project and calls it a decision. Real decisions have an alternative, and the writer can almost always name it without thinking.
The number. I'd look for an owner's relationship to the figure. A strong answer says where the number started, roughly where it landed, and what the person did that moved it, including the part that was luck or somebody else's work. A thin answer repeats a percentage from the appraisal with nothing underneath it. If you have the real figure from your own dashboard, hold it back and compare quietly. Do not turn it into a quiz in the meeting.
What they would do differently. This is the one a generic summary struggles with most, because it needs the writer to have actually looked back. A strong answer is small and slightly uncomfortable: "I'd have raised the staffing problem in March, not June." A thin answer is a dressed-up strength, such as "I'd probably try to delegate more, because I care so much about quality." Take the small and slightly uncomfortable answer every time. It is the best evidence of self-awareness a rating form will ever see.
Reading the answers fairly
Some people will write the follow-up in their own plain, slightly rough voice, and it will sound nothing like their appraisal. That is fine, and it is often a good sign. You are not looking for a match in style. You are looking for facts that hold together and that you recognize from the year.
Some people will answer thinly, and that is useful to know, but I'd treat it as a prompt for a question rather than a verdict. Quiet people, new hires, and anyone writing in a second language may need a day and a nudge. In the 1:1 you can say, "Walk me through the decision one more time, out loud," and listen. Most people who did the work can talk about it for two minutes without notes.
I'd also keep your own file. Three or four dated lines per person, kept through the year, would have settled most of this already. If you have none, start this week for next year. The follow-up questions are a repair for a year that went by without notes, and there is no shame in using one.
One practical note: look at what your company's policy says about using AI tools to prepare appraisals before you comment on it. Many have not written a line yet. If yours has, follow it. If it has not, I would not make up a rule in the middle of review season and enforce it on eleven people.
Opening the 1:1
You do not need to mention the tool at all. I'd open with something like this:
"I read your appraisal and your three answers. The part I'd like to spend our time on is the decision you described. Tell me what you were weighing."
That sentence tells the person that evidence matters here, and that the conversation is about their work. Whoever wrote the first draft, the person across the table now has to talk about what they actually did.
What this gets you
By the end of next week you will have rated eleven people on what they did rather than on how smoothly they described it. Your ratings will be easier to defend when your own boss asks "why a four?", because each one rests on a decision, a number and a lesson you can name. The polished-but-empty appraisals will have sorted themselves out without anyone being accused of anything.
Send the message on Monday morning, then close the shared folder. Eleven very capable-sounding strangers can wait until the answers come in.
P.S.
If you'd like more questions that get specifics out of people without feeling like an interrogation, The AI Meeting Survival Kit puts them on one page. It also has a flip card of the AI terms that show up in leadership rooms. ($19. The three questions above work on their own.)
One small favour. If The AI Jargon Buster has helped you in a meeting, an honest review on Amazon or Goodreads helps the next manager find it. Two lines is plenty. Good, bad or "it's fine": all of it helps. Review on Amazon · Rate on Goodreads
This was a free one. There's a new one every week.
What's signal in the AI noise — and the move to make about it. No hype, no vendor pitch, no link dump.
Aiden Vector is an AI-assisted publication; this content is produced by AI under human editorial direction.