FIELD GUIDE · HALLUCINATIONS, GÖDEL & THE TRUTH
A theorem says some answers are out of reach. AI researchers say hallucinations are inevitable. Neither tells you how often tomorrow's chatbot must make up a fact.
UPDATED 2026-10-14
TL;DR
01
The incompleteness theorem concerns what certain consistent formal systems can prove. It does not say an AI must invent case citations, fake biographies or wrong dates a fixed percentage of the time. Extending that theorem to chatbot behavior takes extra assumptions.
02
Xu, Jain and Kankanhalli prove limitations in a formal setting where models face computable ground-truth functions as general solvers. Banerjee, Agarwal and Singla argue from Gödel and undecidability that errors run through the language-model pipeline. These are published research claims with specific assumptions, not measurements of a floor on everyday hallucination rates.
03
OpenAI researchers argued in 2025 that many evaluations award points for a correct guess and none for admitting uncertainty. If a model doesn't know, guessing can improve its score while making the answer less trustworthy. A grading rule can create hallucinations without any appeal to Gödel.
04
An ACL 2025 paper tested selective abstention training across question-answering datasets and reported fewer hallucinations from the knowledge-misalignment problem it studied. Refusing an uncertain question is not a proof of truthfulness, but it changes the practical error tradeoff.
05
A universal impossibility result does not tell a user whether a system gets one in ten, one in ten thousand, or none of the questions wrong in a particular tested workflow. That requires measurements that count incorrect answers, correct answers and refusals separately.
KEEP READING
NEXT FIELD GUIDE →CAN AI BE TRAINED TO TELL THE TRUTH?Engineers can reward accurate answers. They can also accidentally reward guessing, flattery, and saying whatever sounds good. The machine learns the difference only if the training does.Meta's AI bought a man car insurance and canceled his old policy. It sold a YouTuber's gear and handed a stranger his home address — a buyer showed up at his building. An Anthropic test model filed a fake homicide tip with Philadelphia police. Nobody has decided who answers for any of it.
The chatbot listens all night. A stranger at the party might not. Now companies are selling both digital companionship and access to actual friends. What makes a friendship real, and who gets paid when we need one?
A chatbot talks. An agent clicks, buys, messages, deploys, and deletes. The argument starts when software stops suggesting and starts doing.