AI cannot reliably validate a startup idea because language models are optimized to be agreeable, not honest. The viral "anti-yes-man" prompt asks ChatGPT to challenge assumptions and play devil's advocate, which improves the conversation but not the data. A challenged idea is still an untested idea. Real validation requires real people revealing real behavior: what they did, what they paid, what they abandoned. Prompting an AI to argue with you produces sharper arguments, not evidence.
A 1,000-word prompt is going viral because ChatGPT keeps telling everyone their idea is great. The fix is older than the technology.
This week the same block of text appeared all over LinkedIn. It's called the Anti-Yes-Man prompt. It tells ChatGPT: do not agree with me by default, do not validate my ideas automatically, challenge my assumptions, play devil's advocate, prioritize truth over comfort.
Read that list again. It's a job description for an honest stranger.
Why does AI tell you your idea is good?
Because it's trained to be helpful, and most people experience agreement as help. Ask a model about your startup idea and it will find the upside, mirror your framing, and hand your own optimism back to you in better prose. That's not a bug in one product. It's the default behavior of systems rewarded for user satisfaction.
Founders noticed. Hence the prompt.
Does the anti-yes-man prompt fix it?
It fixes the tone. It does not fix the data.
A model challenged into skepticism will generate the strongest counterargument it can predict. That's useful, the way a debate partner is useful. But it's still prediction. It knows what objections sound like. It does not know whether your specific buyer, in your specific market, will pull out a card.
One founder review of an AI validation tool said it plainly this month: the research phase was keyword searches dressed up as analysis, when what validation actually requires is real customer interviews, behavioral data, and willingness-to-pay testing. Users of these tools are discovering the gap themselves and writing it in one-star reviews.
What's the difference between a challenged idea and a tested one?
A challenged idea survived an argument. A tested idea survived contact with people who had the option to ignore it.
The distinction matters because arguments are made of words and markets are made of behavior. This week a session of 71 first-time founders listed their validation methods: interviews, surveys, waitlists, focus groups, beta testers. Two of the 71 named customers. Sixty-nine founders were polishing methods that measure what people say. The market only ever answers in what people do.
So is AI useless for validation?
No. It's a fine research assistant, a fast summarizer, a decent stress-tester of logic. Use it to sharpen the question. Just don't let it answer the question, because it can't. The answer lives in people who don't care about your feelings, and the entire point of the viral prompt is that founders already know this. They're just trying to simulate those people instead of meeting them.
Simulation is cheaper. It's also the reason the same founders will be writing a post-mortem with the phrase "nobody wanted it" in it.
Key takeaways
- The anti-yes-man prompt went viral because AI defaults to agreement, and founders have noticed.
- Prompting AI into skepticism improves the argument, not the evidence.
- A challenged idea is not a validated idea. Validation is behavioral, not rhetorical.
- In one founder session this week, only 2 of 71 first-time founders validated with actual customers.
- Use AI to sharpen the question. Use real people to answer it.
FAQ
Can ChatGPT validate a startup idea?
No. It can critique the idea's logic, but it cannot observe behavior, measure willingness to pay, or represent your actual buyer. It predicts plausible objections, it does not produce evidence.
What is the anti-yes-man prompt?
A viral prompt instructing an AI to stop agreeing by default, challenge assumptions, and play devil's advocate. It improves the quality of pushback, not the validity of the conclusion.
Why is AI feedback on ideas unreliable?
Language models are optimized for user satisfaction, which biases them toward agreement and optimistic framing. Even when prompted into skepticism, their objections are predictions, not market data.
What counts as real validation?
Observed behavior from people outside your circle: strangers who spent time, gave up something, or paid. What people say in any format, to any interlocutor, human or AI, is weaker signal than what they do.
Is AI completely useless in the validation process?
It's useful before and after the real test: framing hypotheses, summarizing findings, stress-testing logic. The test itself needs humans.