A hallucination is when an AI model states false or invented information as if it were fact — a citation that doesn’t exist, an API that was never shipped, a confident wrong answer. The defining feature isn’t the error itself but the fluency: hallucinated output reads exactly like correct output, which is what makes it dangerous.
Why models make things up
An LLM generates by predicting plausible next tokens, not by consulting a fact database — there is no internal “true/false” check between the model and its output. When the training data is thin on a topic, the most statistically plausible continuation can simply be wrong. Training incentives make it worse: models rewarded for giving answers learn that a confident guess scores better than “I don’t know” — OpenAI’s 2025 research framed hallucination as exactly this test-taking behavior. Errors also compound in long agent runs: Anthropic’s Fable 5 system card documents the failure mode of an agent fabricating a status report for work it silently failed to do.
What actually reduces it
- Grounding (RAG): retrieve real documents into the context window and instruct the model to answer from them — models are far more accurate about text in front of them than facts recalled from training.
- Tools: let the model search, run code or query databases instead of answering from memory.
- Calibration: newer frontier models are explicitly trained to flag uncertainty rather than bluff — Claude Opus 4.8’s main selling point is exactly this, and it’s why calibrated models are preferred for legal, finance and autonomous-agent work.
- Verification: for anything high-stakes, check citations and keep a human approval gate on irreversible actions.
Mitigated, not solved — treat every unverified factual claim from a model as a draft. See how calibration shapes real model choices in the best LLMs for agents ranking.