Why ChatGPT says your startup idea is good (and how to get a real answer)

The mechanism, briefly

Assistants like ChatGPT are tuned using human feedback, and human raters tend to prefer responses that are agreeable, encouraging, and responsive to what was asked. Over many rounds that preference becomes a habit: the model leans toward the answer that fits the frame it was handed.

Ask "is my idea for X good?" and you have already supplied the frame. The question presupposes the idea is the thing to evaluate, and the cooperative completion is a balanced-sounding answer that leads with strengths. You will usually get caveats too, but they arrive as a polite second half rather than a verdict.

The bias is not that the model lies. It is that you set the frame, and the model works inside it. Change the frame and you change the answer.

Three tells that you are getting a cooperative answer

• It never names a specific competitor, or names only famous ones that anyone could list from memory.
• The risks section is generic: execution risk, competition, customer acquisition. These apply to every idea ever conceived and therefore say nothing about yours.
• It ends by asking whether you would like help with a go-to-market plan. The conversation has quietly moved past the decision without ever making it.

Prompts that get a real answer

You can get genuinely useful critique out of a general assistant. It just requires deliberately removing your own frame.

Things that work, roughly in order of effectiveness:

• Strip your ownership. Paste the idea as "a founder I am advising is proposing this" rather than "my idea." Removing your visible investment removes the pull toward encouragement.
• Ask for the failure post-mortem. "It is two years later and this failed. Write the honest retrospective explaining why." This inverts the frame entirely and tends to surface the real risks.
• Force a decision, not an essay. "Answer build or do not build in the first line, then justify it." A model that must commit before elaborating cannot hedge its way through.
• Demand named specifics with sources. "Name the five companies already doing this, with links. If you cannot find five, say so explicitly rather than filling the gap." This is also how you catch invented competitors.
• Ask what would change your mind. "What single piece of evidence would make this clearly not worth building?" Then go and look for that evidence.

Run these in a fresh conversation. Once a thread has been enthusiastic for several turns, that enthusiasm is context the model keeps reading.

The limit you cannot prompt your way around

Even with good prompting, two problems persist. The first is consistency: the depth of the answer depends on how you phrased the request and whether the model chose to search, so two founders asking about the same idea can get very different quality.

The second is harder. If a model is answering partly from training data rather than from live retrieval, it can name a company that fits the shape of your market without having verified that company is currently in it, or still trading. Well-known names are the most likely to appear this way, because they are the most strongly represented in the training data.

Both are structural. They are not evidence the model is bad, only that a general assistant is being asked to do a specific job, and the quality of that job is left to you.

What a purpose-built tool changes

Founder Evolution AI is built around removing exactly these two failure modes. It runs the same research sequence for every idea, so depth does not depend on prompting skill. It decides the verdict type in code from what the research actually found rather than letting the language model choose it, which removes the part most likely to drift toward agreement. And it can only name companies that appeared verbatim in live search results it retrieved, so it cannot fill a gap with a brand from memory.

None of that makes a general assistant useless here. ChatGPT is genuinely better for thinking out loud, reshaping a rough concept, and planning execution once you have decided to build. It is the go or no-go moment specifically where the cooperative instinct works against you.