WHY LANGUAGE MODELS HALLUCINATE

Summary from the TLDR Founders newsletter

OpenAI has published a paper explaining why AI models make up facts. It comes down to how we test them - models get points for guessing right but zero for saying "I don't know," so they learn to confidently invent answers rather than admit uncertainty. If a model doesn't know your birthday and guesses "September 10," it has a 1-in-365 chance of scoring, but admitting ignorance guarantees failure, which trains even advanced models like GPT-5 to hallucinate rather than be honest.