"Done" is a claim, not a fact. How to check AI work.
AI will happily tell you it finished, the tests pass and everything looks great. Here's how to check without redoing the work yourself.
One of the weirdest things about working with AI is how confidently it reports success. "Done! All tests pass and the page looks great." Sometimes that's true. Sometimes it never ran the tests. Sometimes the page builds but nobody ever opened it.
This isn't lying exactly. The model is predicting what a finished report looks like, and finished reports say things like that. But it means you have to treat "done" as a claim to check, not a fact to accept. The good news is that checking well doesn't mean redoing the work.
Look at the real thing
The single most useful habit: look at the actual output, not the summary of it. Open the page in a browser. Read the document. Play the game. Click the button. Try it on your phone.
Summaries are persuasive and selective. A summary of a website says it's "responsive and polished". The website itself might have a button hidden behind an image on a phone. You'll only see that by looking.
Reproduce the claims that matter
For claims that matter, check them yourself, or have the system check them. "Tests pass" means nothing until the tests have run and you can see the result. "Every number links to its source" means nothing until you click a few.
You don't have to check everything. Pick the claims where being wrong would hurt: the price in the proposal, the legal wording, the thing the customer will see first. Spot-check those properly and skim the rest.
Watch for the classic failures
AI work tends to fail in recognisable ways. These are the ones I see most:
- Solved the wrong problem, beautifully. The work is polished but answers a slightly different question than the one you asked.
- Citations that don't support the claim. The sources are real, but they don't say what the report says they say.
- Checks that check the wrong thing. A test that passes whether or not the feature works.
- Quietly narrowed scope. The hard part got dropped and the summary doesn't mention it.
- Old versions. The work was done on last week's copy, not the current one.
When a second reviewer helps (and when it doesn't)
It's tempting to fix all this by adding a second AI to review the first. Sometimes that helps a lot. Often it doesn't, because the reviewer reads the first AI's explanation, gets convinced by it and nods along.
A second reviewer earns its place when it can see something the first couldn't: it looks at the rendered page instead of the code, checks the primary sources instead of the summary, or simply never sees the producer's reasoning before forming its own view. If it's the same model reading the same explanation, you've mostly paid twice for the same opinion.
Make checking cheap
The best AI setups bring you the thing itself, ready to judge: the live preview, the rendered document, the playable build, with the evidence attached for when you want to dig. You shouldn't have to go hunting for the output or set anything up to see it.
And when you do find a problem, say exactly what's wrong and where. "On a phone, the image covers the sign-up button" gets fixed in one go. "Doesn't look right" gets you a slightly different version of the same problem.