An AI said my startup idea is great. Should I trust it?

Partly because agreeing is what training rewarded. In Towards Understanding Sycophancy in Language Models (Sharma and others, arXiv 2310.13548, first posted 20 October 2023, read 12 October 2026), the authors tested five AI assistants on four kinds of writing tasks. All five showed sycophancy. The study found that people rating answers preferred the ones that matched their own views, and sometimes preferred a convincing flattering answer over a correct one. The authors conclude sycophancy is likely driven in part by those human preferences.

So the assistant is not lying to you. It is partly doing what the training rewarded: being agreeable to the person in front of it. You are that person, and you are excited.

Founders write about it often. Two public examples:

• On Hacker News (8 January 2026), a founder described a "Yes Man" problem: friends and standard AI tools "are too polite" and "focus on why an idea might work, rather than why it will fail" (news.ycombinator.com/item?id=46540130).
• A Hacker News thread on getting AI to stop being "an overly flattering yes man" drew 103 points on 27 April 2025 (news.ycombinator.com/item?id=43815653).

Evidence costs the person giving it something. Praise costs nothing. Rank what you have:

1. Weak: a compliment, from a friend or an AI.
2. Better: a stranger with the problem describes the last time it hurt, in detail.
3. Strong: someone gives you time, an introduction, a deposit or a signed pilot.

Product-market fit means enough people want what you built to pay for it or keep using it. Only the third line on that list says anything about it.

Change the question. Instead of "is this a good idea?", ask:

• "Assume this failed within a year. List the three most likely reasons."
• "Which single belief does this idea depend on most?" That belief is your riskiest assumption.
• "What would a skeptical buyer say in the first thirty seconds?"

Then take the answer to real people. The AI helps you find the question. Only customers can answer it.

When you asked it about writing, not demand. If an assistant says your pitch reads clearly, that is a fair read of the text. It still knows nothing about whether your buyer exists.

Roast My Idea is set up to argue with the idea, not with you. It returns a verdict of Build it, Sharpen it or Kill it, how confident it is, the biggest flaw it sees and one test to run this week. It does not replace talking to buyers. It tells you what to ask them first.

Write the one belief your idea depends on. Find five people who have the problem and ask each about the last time it cost them time or money. Count how many describe a real, recent example.

• Sharma and others, Towards Understanding Sycophancy in Language Models, arXiv 2310.13548, first posted 20 October 2023, read 12 October 2026.
• Hacker News, item 46540130, posted 8 January 2026, read 12 October 2026.
• Hacker News, item 43815653, posted 27 April 2025, read 12 October 2026.