How to Fact-Check an AI-Written Performance Review Before You Submit It
AI can draft a solid review in seconds, but it can also invent facts and mask bias. Here's how to check the work before you hit submit.
How to Fact Check an AI Written Performance Review Before You Submit It
If you're using AI to help write performance reviews, you're not doing anything wrong. You're doing what smart managers and employees alike do when they're short on time and staring down a blank self assessment or peer review form. AI assisted reviews can be genuinely great when they're built the right way, and I'll say more about that later.
But here's the catch nobody talks about. Generative AI doesn't retrieve facts. It predicts what sounds plausible. That's a huge difference when the document in question decides someone's raise, promotion, or continued employment.
So before that AI drafted review goes anywhere near your employee, it needs a proper fact check. Not a skim. A real one.
Why This Actually Matters
AI models are known to hallucinate. That's the polite industry term for confidently making things up. A tool might invent a project that never happened, misstate a metric, or turn one offhand comment into a sweeping character judgment.
This isn't a fringe concern either. Even the companies building these models track hallucination rates as a core metric, because they know their own tools can be wrong while sounding completely sure of themselves.
Performance reviews aren't a low stakes place for that kind of error. Bodies like the CIPD and regulators such as the EEOC have been clear that the responsibility for fair, accurate outcomes sits with the employer, not the AI vendor. If a review contains a fabricated detail or biased phrasing, that's on you and your organization, not the model that generated it.
Where AI Reviews Tend to Go Wrong
Made up facts and dates
AI tools can invent specific projects, numbers, or incidents, especially when the prompt is vague. They can also take a single piece of feedback and blow it up into a permanent character trait, like turning one missed deadline into "consistently struggles with time management."
Bias hiding in plain sight
Performance reviews have always been vulnerable to bias. Recency bias, similarity bias, and yes, plenty of gendered language too. AI doesn't fix this. It's trained on historical data, which means it can just as easily repeat old patterns as avoid them. Regulators have flagged that algorithmic tools used in employment decisions can create unlawful disparate impact if nobody is checking for it.
Generic voice, missing context
AI writing has a certain flavor to it. Smooth, competent, and a little hollow. It often misses the specific team norms, values, or context that make a review feel like it was actually written by someone who knows the employee. When that happens, feedback can come across as insincere even when the substance is fine.
The Core Principle: Draft, Not Decision
AI governance guidance keeps coming back to one idea. Treat AI output as a draft, never a finished decision. The organization and the manager remain accountable for what ends up in that document, not the tool that produced it.
This is really the mindset shift that matters most. AI is there to help you articulate things faster, not to replace your judgment about what actually happened and what it means.
A Practical Fact Checking Workflow
1. Get your evidence together first
Before you even open the AI tool, gather your actual evidence. Objectives, KPIs, 1:1 notes, project outcomes, any 360 feedback. Make sure this evidence itself is accurate before it becomes the basis for anything else.
2. Prompt it properly
Tell the AI explicitly to only use the facts you provide and not to invent details. Ask it to separate factual statements from evaluative ones. This single step saves enormous amounts of fact checking time later, because you're not untangling opinion from fact after the fact.
3. Check the shape before the details
Does the draft cover the whole review period, or is it weirdly focused on the last few weeks? Are strengths and development areas both represented? If the structure is off, fix that before you start checking individual sentences.
4. Go line by line
This is the real work. For every factual claim, ask yourself where the evidence is. Highlight every mention of a project, a metric, a date, or a specific behavior. If you can't point to a source, cut it or rewrite it. Numbers and names are the most common places hallucinations sneak in, so give those extra scrutiny.
5. Check for consistency
Compare the review against similar reviews on your team. Are you seeing noticeably harsher or more glowing language for particular people without a clear reason? Inconsistent standards across comparable employees are a fairness problem waiting to happen.
6. Hunt for biased language
Words like "assertive" and "aggressive" often describe the exact same behavior depending on who's being described. Same goes for "detail oriented" versus "pedantic." Read the draft specifically looking for this kind of double standard.
7. Match it to policy and law
Check the review against your organization's performance framework and any relevant employment law in your jurisdiction. This is the boring step nobody enjoys, but it's the one that protects everyone if a decision is ever questioned.
8. Separate fact from judgment
There's a real difference between "delivered the release two weeks early with zero critical incidents" and "showed strong ownership this quarter." The first is verifiable. The second is your interpretation. Good reviews make that distinction clear rather than blurring the two together, which AI tends to do by default.
9. Write down what you checked
A quick note that AI was used as a drafting aid, plus confirmation that claims were checked against real sources, goes a long way. It doesn't need to be elaborate. It just needs to exist.
10. Get a second set of eyes
A short calibration conversation with HR or a peer manager can catch things you missed, especially patterns that only become visible when you're comparing multiple reviews at once.
A Quick Checklist Before You Hit Submit
Evidence gathered and verified AI prompted to stick to supplied facts only Draft covers the full period with balanced strengths and development areas Every factual claim checked against a real source Language consistent with how you've described comparable employees No stereotyped or coded wording Aligned with your performance framework and legal obligations Verification steps noted somewhere
The Bottom Line
I actually think AI assisted performance reviews are a good thing when they're done properly. They save time, they help with structure, and they take some of the awkwardness out of finding the right words. The problem was never the AI itself. It's skipping the verification step that turns a helpful draft into a liability.
Treat the AI like a capable junior writer who's occasionally confidently wrong. Check its work, keep your own judgment in the driver's seat, and the reviews you send out will be both faster to produce and genuinely trustworthy.
How Perform Review Can Help
If you want AI assistance without the guesswork, Perform Review is built specifically for high quality self, peer and manager assessments. It helps you draft thoughtful, evidence based reviews with AI support while keeping you firmly in control of accuracy and tone, so what you submit is something you can stand behind.