The work comes in on time. It looks finished. The formatting is clean, the structure is there, the sentences are smooth. And something is wrong with it. A statistic that doesn't have a source. A "client quote" nobody can place. An analysis that's confidently structured around a conclusion the data doesn't actually support. You can't always point to the exact failure, but you've stopped trusting the work, and you're now re-checking everything they hand you, which means you're doing the job twice.
You're pretty sure they're using AI. And your instinct is to prove it.
Drop that instinct. It's the wrong fight, and you'll lose it. The moment you say "did you use ChatGPT for this," you've turned a performance conversation into an interrogation about tools, and they will simply get better at hiding it. You'll be playing forensic detective on every deliverable instead of managing output. And the secrecy is structural, not personal: a 2025 Laserfiche survey found 49% of employees who use AI at work keep it to themselves, and one in seven actively avoid telling their manager — usually out of fear of looking lazy or replaceable. You will not interrogate your way out of a problem half the workforce is already concealing.
Here's the reframe. The tool isn't the problem. The problem is that you're accepting work below your standard, and you never made the standard explicit because you never had to before. AI didn't lower this person's bar. It made visible a bar you'd been enforcing implicitly through effort — "they clearly worked hard on this, so I'll trust it." That heuristic is now broken. It's not coming back. So manage the only thing that was ever actually your job: the output.
Three moves.
First, raise the visible bar and make it about accountability, not provenance. Don't say "I think you used AI." Say: "There were two fabricated quotes in this and an unsupported claim in the analysis. I need you to be able to defend every fact and every conclusion in anything you hand me. If you can't source it, it doesn't go in." Now the standard is explicit, it's about the work, and it doesn't matter how the work was made. AI use is fine. Unverified output is not. That's a defensible line and a true one.
Second, make them present, not just submit. The fastest way to surface hollow work is to stop accepting it asynchronously. Have them walk you through the reasoning live — "talk me through how you got to this recommendation." Real work survives that conversation. Hollow work falls apart in about ninety seconds, and they learn that "looks done" isn't the same as "is done" — without you ever having said the word AI.
Third, separate the two failures, because there are two. One is a quality problem: they're shipping work they haven't checked. The other is a judgment problem: they couldn't tell the difference between solid work and confident garbage. The first is fixable with a clear bar. The second is more serious — it tells you whether this person actually understands the work well enough to do the job at all. Diagnose which one you have. They require different conversations.
A caveat: don't overcorrect into banning the tool. The people doing this well are using the same software — they're just verifying the output. If you make AI itself the offense, you punish your good people for being honest about their process and teach everyone else to hide it better. The behavior you're correcting is shipping unverified work, not using a tool.
The uncomfortable part: you can't actually tell who's using AI, and you never reliably will. What you can tell is whose work holds up when you press on it. That was always the only thing you could measure. You just never had to until now.