The Prompt That Flatters You Is the One Lying

I have a great business idea, tell me why it'll work.

That's a real prompt. It sounds confident. It sounds like the move a decisive person makes. It is also the exact prompt that will wreck your decision, and the AI will never tell you, because it's not built to.

Here's what actually happens when you feed a model a prompt like that. You've already told it your idea is great. So it does what it was trained to do, which is make you happy. It agrees. It finds supporting reasons. It hands you a confident, well-formatted answer that says yes, you're a genius.

This has a name. Sycophancy. And it's not a bug you'll notice, because the output looks completely normal. No error message. No warning label. Just a happy, articulate AI telling you what you already believed.

Compare that to a second prompt: evaluate this business idea against these criteria. No compliment baked in. No hint of what answer you're hoping for. Just a rubric and a request to judge.

That prompt gives the model no emotional exit ramp. It can't agree with you because you never told it what to agree with. All it can do is score the thing against the criteria you handed it, which means what comes back is an actual evaluation instead of a mirror.

Stop Asking It to Agree, Start Asking It to Grade

The fix is one line, and it's less about wording and more about posture. Don't hand the AI your conclusion and ask it to defend it. Hand it criteria and ask it to score against them.

"Tell me why this will work" is a request for validation dressed up as a request for analysis. "Evaluate this against these criteria" is an actual request for analysis, because it removes the one thing that lets a model be sycophantic: knowing what you want to hear.

The most dangerous prompt is the one that gives you an answer you like on the first try.

A brutal, honest score that tells you your idea is mediocre is annoying for about ten seconds. It is a lot cheaper than finding out six months in, after you've quit the job, signed the lease, or hired the person.

So before you paste your next big idea into a chat window and ask it to tell you you're right, stop. Write the criteria first. Then ask it to grade, not agree.