For Developers/Glossary/Hallucination
Evaluation

Hallucination

When a language model states something false or invented, such as a fact, a citation or an action it never took, with the same fluency and confidence as accurate output.

LLMs don't look facts up in a database. They generate text from learned statistical patterns, and sometimes those patterns produce false statements with the same confident fluency as true ones. Without a tool or source to check against, the model has no reliable way to tell a well-grounded answer from a plausible guess.

Types of hallucination

  • Factual hallucination: Inventing specific facts (dates, statistics, names) that don't exist or misrepresenting known facts.
  • Citation fabrication: Generating plausible-sounding paper titles, DOIs, URLs, or author attributions that don't exist.
  • Action hallucination: Claiming to have performed an action (reading a file, running code, searching the web) that wasn't actually done.
  • Temporal confusion: Stating outdated information as current fact, or treating the training data cutoff as the present day.

Why models hallucinate

Pre-training rewards predicting plausible text, so on rare or missing topics a model produces something that sounds right instead of declining. OpenAI's September 2025 paper "Why language models hallucinate" adds that training and evaluation reward guessing: on an accuracy-only benchmark a guess sometimes scores, while "I don't know" never does. OpenAI argues that benchmarks should penalize confident errors more than expressions of uncertainty.

Measuring hallucination

Independent evaluators now track hallucination as its own metric, and results depend heavily on the task:

  • Knowledge questions: Artificial Analysis's AA-Omniscience asks hard knowledge questions and defines the hallucination rate as the share of wrong answers among all responses that were not correct, so declining to answer helps a model's rate. In its September 30, 2026 launch analysis, Gemini 4 Argon (high) hallucinated on 15% of those questions against 51% for GPT-6 Astra (max), but answered 50% correctly against Astra's 63%. Version 1.1 of the benchmark (October 31, 2026) removed saturated questions, so its scores are not comparable with the first version. It lists Argon (high) at a 4% hallucination rate. On accuracy, Claude Fable 5.2 (max effort) took the top spot at 79% after its November 9 release, ahead of the earlier leaders Claude Fable 5.1 (68%) and Claude Opus 5.5 (66%).
  • Summarizing a supplied document: Vectara's hallucination leaderboard (updated October 22, 2026) checks whether summaries add claims the source does not support. Here the order flips: GPT-6 Astra added unsupported content in 8.7% of summaries, Claude Opus 5.5 in 9.9% and Gemini 4 Argon in 12.1%, and the lowest rates on the board come from smaller models such as GPT-5.4 nano (3.1%).
  • Vendor-reported tests: OpenAI says GPT-6.1 Sol cut its factual error rate to 7.7% from GPT-6 Sol's 11.4% at low reasoning effort, measured on de-identified conversations where users had flagged an earlier model's error. OpenAI notes that these prompts are harder than typical use.

Read a hallucination rate next to accuracy and the type of task. A model can lower its rate by declining more often. That suits legal and finance work, where a confident wrong answer is costly, while open-ended research may favor a model that attempts more answers. See our Gemini 4 Argon analysis for the launch numbers.

Reducing hallucinations

  • Retrieval-Augmented Generation (RAG): Give the model the documents it should answer from, so it does not rely on memory for facts it never saw.
  • Citations and grounding: Anthropic's Citations feature, supported on all active Claude models, returns the exact passages behind each claim, and the API guarantees that each citation points to a real location in the documents you supplied. Google's Grounding with Google Search, available on Gemini 4 Argon and on Gemini 4 Hydrogen's paid tier, attaches URL citations to spans of the answer. A citation shows where a claim came from, not that the source supports it, so spot-check the ones that matter.
  • Tool use: Let the model call search, databases or code execution instead of producing facts from its weights.
  • Constrained output: Request JSON with required fields or other structured formats. Constraints reduce but don't eliminate hallucinations.
  • Allow abstention: Tell the model that "I don't know" is an acceptable answer and ask it to flag low-confidence claims. Model-reported confidence does not reliably match accuracy, so verify claims that matter.
  • Self-consistency: Generate multiple completions and compare answers. More useful for reasoning tasks than for factual recall.

Related terms

Models relevant to Hallucination