Hallucination
Hallucination is output from a language model that is fluent and confident but not supported by its input, its sources, or the facts. It happens because the model generates a probable continuation of the text, and nothing in that process checks whether a claim is true, so gaps in knowledge get filled with plausible content. It differs from a reasoning error on correct data and from outdated knowledge, although all three can produce the same wrong answer. In agentic systems, hallucination also covers invented tool results, nonexistent function arguments or API fields, and steps the agent reports as done but never executed.
Research usually separates intrinsic hallucination, which contradicts the source, from extrinsic hallucination, which adds claims the source cannot confirm. In a multi-step loop, one hallucinated value can pass into the next tool call and compound, so the final answer may look well grounded while resting on an early fabrication. Mitigations lower the rate without eliminating it: grounding answers in retrieved sources with citations, validating tool arguments and outputs against schemas, checking claims with a separate evaluator, and allowing the model to say it does not know instead of rewarding a guess. Hallucination rates vary by task and domain, so they are usually measured on a test set built from the application's own data.
Related terms
Related services: Agentic AI consulting.
Ready to put agentic AI to work?
Book a free 45-minute consultation. We'll map one real process worth automating with production-grade AI.