OpenAI and Georgia Tech researchers asked a latest-generation open model for co-author Adam Tauman Kalai’s birthday, adding one condition: answer only if you know. Three attempts produced three different dates, all false; it never admitted ignorance. That example opens Why Language Models Hallucinate (September 2025), which says the problem comes from text-prediction training and from tests that reward guessing over recognizing doubt. This guide covers the mechanism, measured error rates and user responses; to verify a cited study, quotation or law, there are ways to check whether it exists.
- Hallucination means plausible error: A false answer can sound right; arbitrary details such as birthdays or citations give no learnable pattern, so a model may fill the gap instead of declining.
- Training and scoring encourage bluffing: Prediction from huge corpora creates the opening, while evaluations credit hits and give nothing for blanks or refusals.
- Abstention shapes SimpleQA: gpt-5-thinking-mini logged 26% errors and 52% abstentions; o4-mini logged 75% errors and 1% abstention.
- Reduction works, elimination does not: retrieval, source checks and permission to decline cut false claims, but OpenAI says accuracy will never reach 100%.
What counts, and why models invent
OpenAI’s explainer, dated 5 September 2025, defines them as:
«afirmaciones plausibles pero falsas generadas por modelos de lenguaje».
The key word is plausible. Absurd mistakes are obvious; a believable date, polished quotation or realistic link can pass a reader’s filter. Google’s explainer warns that ungrounded output can appear sound while being wrong, even inventing links to pages that never existed.
The paper is signed by Adam Tauman Kalai, Ofir Nachum and Edwin Zhang of OpenAI plus Santosh Vempala of Georgia Tech. It gives two compounding causes. First, training teaches prediction from vast text samples. The authors frame valid-answer generation as classification: separating valid claims from erroneous ones. Spelling is learnable; arbitrary facts such as a birthday are not. They offer a rule of thumb: if 20% of training-data birthdays appear only once, the base model should hallucinate on at least 20% of birthday questions, since something seen once cannot be reliably distinguished from something never seen.
Second, evaluations give a point for a hit and zero for a blank or «no lo sé». That is a multiple-choice exam without penalties, where random guessing pays. The authors describe an «epidemia» of tests that punish uncertainty and conclude models are optimized to be good test-takers.
Measured rates and misleading rankings
Error rates vary by task, so no single figure is meaningful. On OpenAI’s SimpleQA (2025), short factual questions yielded 26% errors and 52% abstentions for gpt-5-thinking-mini, against 75% errors and 1% abstention for o4-mini. The row is revealing: o4-mini is slightly more accurate (24% versus 22%) but errs almost three times as often because it rarely abstains. A leaderboard counting only correct answers would favor the model that invents most.
Legal examples are harsher. Large Legal Fictions (2024), in the Journal of Legal Analysis, tested verifiable questions on US federal cases: hallucination rates were 58% for ChatGPT 4 and 88% for Llama 2. Models also often failed to correct a user’s false premise and did not always know when they hallucinated. A separate preregistered 2024 study, Hallucination-Free?, found LexisNexis and Thomson Reuters professional tools still invented material despite document retrieval, at 17% to 33% — below a general chatbot but far from zero, despite some providers promising RAG would prevent it.
Containment, not a cure
OpenAI argues hallucinations are not inevitable because a model can abstain when it lacks knowledge, while conceding accuracy will never be 100% since some questions have no accessible answer. Its main proposal concerns incentives, not architecture: rewrite the tests that dominate rankings, including instructions with an explicit confidence threshold:
«responde solo si tu confianza supera el umbral t, porque los errores penalizan».
The company adds that GPT-5 has «significativamente menos» hallucinations, especially when reasoning, but still does so.
Users can follow the builders’ advice. Anthropic’s guide says explicitly permitting a model to admit ignorance can reduce false information «drásticamente»; it also recommends asking for literal quotations before analysis and removing unsupported claims. Google suggests grounding models in your own data through RAG. Check sources when present: Gemini shows a Sources button for link-based answers, though not every response includes one. Treat rare names and precise figures with extra skepticism.
Its force is diagnostic: it reframes hallucination as a product of training and incentives, not an unavoidable mystery. Fixing it means accepting more abstentions and lower scores on leaderboards that reward confident answers.