Hallucination

A confident but incorrect output from an LLM — invented facts, fabricated citations, or nonexistent functions — produced as if it were grounded.

Hallucination is the defining failure mode of LLMs. The model generates plausible-sounding text that isn’t true — wrong dates, invented book titles, citations to papers that don’t exist, function signatures with parameters the API doesn’t have. The model isn’t lying; it’s pattern-matching on training data without a ground-truth check.

For agents, hallucination has additional surfaces. Hallucinated tool calls (the model invents a function name that doesn’t exist), hallucinated outputs (the agent reports completing a task it didn’t), and hallucinated facts inside otherwise-correct outputs.

Mitigation in 2026 centers on grounding. RAG anchors answers in retrieved documents. Tool calling validates against actual API schemas. Reasoning models (Claude with extended thinking, GPT o-series) hallucinate less than non-reasoning models on complex tasks. Verification steps — separate calls that check the primary output — catch many remaining cases.

Hallucination cannot be fully eliminated. Production agent design assumes some rate of hallucination and builds guardrails and evaluation around the consequences. The Stanford AI Index 2026 reports significant per-model variation; the gap matters when picking models for high-stakes use cases.