Why most AI idea validators say yes, and how to read one that does not

Paste almost any idea into an AI validator and something warm comes back: a score in the seventies, a list of strengths, a market size in the billions, a few "risks to consider" that read like footnotes. It feels like validation. It is a mirror.

Why the yes is built in

Three things push these tools towards approval.

The incentive. A validator that tells a founder their idea is weak loses that founder. One that produces a glowing 100-page report feels like value for money. Over thousands of users, the product that flatters wins, so most products flatter.

The prompt. Ask a language model "evaluate this idea" and it does what it was trained to do with evaluations: balance strengths and weaknesses, hedge, and end on an encouraging note. Unless the prompt forces it to score first and lead with failure modes, prose comes first and the score follows the prose.

The missing data. Most tools work from the model's memory. Memory does not know how many car detailers there are in north Austin or what the three nearest ones charge. So the competition section becomes "the market has several established players", which is true of every market and tells you nothing.

What a real check looks like

A check that can say no has visible machinery:

• Fixed dimensions with weights, published, so a score of 61 means the same thing for every idea.
• Scores before prose. The numbers are produced first; the explanation is written to justify them, not the other way round.
• A verdict from the numbers, computed, not chosen. If the text cannot change the verdict, the model cannot talk an idea up.
• Evidence or "unverified". Every claim about competitors, demand or cost either cites a source you can open, or is marked as unverified. A report with no unverified items is suspicious: every real market has gaps in the data.
• Red flags first. The reasons it fails are the most useful part of any report; they should be at the top and specific to the idea, the market and the place.
• A public distribution. If a tool never publishes how many ideas it approves, assume it approves most of them.

How to read a check that says no

A Kill is not an insult; it is a saving. Read it like this:

1. Find the lowest dimension. That is the structural problem. If it is competition or problem severity, the idea itself is the issue. If it is feasibility or founder fit, the idea may be fine and the plan or the team is not.
2. Read the kill criteria. They tell you the cheap test that could overturn the verdict. If you believe the check is wrong, run that test. Evidence beats both your gut and ours.
3. Look at "what would change the verdict". Sometimes a single fact, a licence that turns out to be easy, a competitor that has closed, moves a Pivot to a Build.
4. Take the pivot seriously. Most ideas that score in the Pivot band have a smaller, sharper version that would score higher. Narrower customer, different price, adjacent problem.

How to use a yes

A Build verdict is permission to spend the next month, not the next year. Run the 30-day plan in the report, honour its kill criteria, and let real customers cast the final vote. The best outcome of any check, ours included, is that you stop reading reports and go talk to ten people.

The methodology page shows our rubric, weights and current verdict distribution. If Build ever climbs above a third, we tighten it.