the no zero hallucinations claim

a founder pitched her ai product last week and used a phrase that made every experienced person in the room flinch.

"zero hallucinations."

she meant well. she had genuinely good rag infrastructure. her grounding work was above average. her false answer rate was lower than competitors.

but zero is a word that has consequences when you use it about ai.

and the room knew it.

this is a small example of a bigger problem. founders overclaim on ai capabilities because the ecosystem rewards bold claims. the press rewards bold claims. the deck templates encourage bold claims. the venture twitter community amplifies bold claims.

and yet.... the actual buyers and the actual investors who matter at scale do not reward bold claims. they reward calibrated claims.

a founder who says "we hallucinate at one third the rate of gpt four on this benchmark" sounds like an engineer.

a founder who says "zero hallucinations" sounds like marketing.

marketing language sounds correct in the press release. it sounds wrong in the conference room.

the calibration problem is everywhere in ai pitches right now.

founders claim their model has reached a benchmark when their dataset was leaked into the training set.

founders claim sub second response time when only their cherrypicked demo runs that fast.

founders claim ten thousand customers when they have ten thousand sign ups and a hundred actual users.

each of these is technically defensible. each of them produces an instant credibility loss in any room with technical investors.

and the credibility loss is permanent. once an investor decides you exaggerate.... they reread every other claim you have made through that lens.

the cost of one overclaim is the discount applied to your entire pitch.

this is why the founders who build the most durable companies often understate. they describe what their product cannot do as well as what it can do. they tell investors where the model fails before being asked.

this counterintuitive move builds trust faster than any feature claim.

because the room understands that you understand. and understanding your own limitations is the highest signal of expertise that exists in any technical field.

nobody who has spent five years on language models believes in zero hallucinations. so when you claim it.... you reveal that either you havent spent the time or you are willing to mislead.

both are bad in the room.

so heres the discipline.

any time you find yourself saying always or never or zero or perfect or one hundred percent.... stop. add a qualifier. give a number. cite a benchmark. show your work.

so what are you currently overclaiming on?