AI hallucination

Large language models are prediction engines, not knowledge bases. They generate text by predicting the most statistically likely next token based on patterns learned during training. They do not "understand" truth; they predict plausibility. When the model encounters a gap in its training data or an ambiguous query, it fills the gap with plausible-sounding fabrication rather than admitting uncertainty.

Researchers distinguish two main types. Intrinsic hallucinations contradict source material the model was given — for example, summarizing a document and inventing a fact that is not in it. Extrinsic hallucinations invent information that cannot be verified against any source — fabricated case law, made-up product specs, fake statistics. Extrinsic hallucinations are harder to detect because there is no original to compare against.

Hallucination rates remain high enough to cause real harm. Stanford RegLab research found that LLMs hallucinate between 69% and 88% of the time on specific legal queries, with at least 75% hallucination rates on questions about a court's core ruling. ECRI, a global healthcare safety nonprofit, listed AI risks as the top health technology hazard for 2025. Courts have imposed sanctions exceeding $10,000 in multiple cases involving AI-fabricated citations submitted by practicing lawyers.

For brands, hallucination affects how AI systems describe your products, attribute your statistics, and represent your positions in generated answers. A confidently wrong AI summary reaches more buyers than an accurate one buried below it.

The Vectara Hughes Hallucination Evaluation Model (HHEM) Leaderboard is the industry's most widely referenced benchmark. It measures grounded hallucination — how often a model introduces false information when summarizing a document it was explicitly given. Vectara provides 1,000+ documents to each model, asks for summaries using only facts from the source, then checks each summary against the original.

Other benchmarks measure extrinsic hallucination by asking models factual questions across diverse domains and checking responses against verified ground truth. Domain-specific evaluations matter: legal hallucination rates run 17% to 34% even on purpose-built legal AI tools, while general-knowledge rates on top models can fall below 1%.

Five techniques have measured impact:

1. Retrieval-augmented generation (RAG). Grounding model responses in retrieved documents at query time reduces hallucination by up to 71% in studies. RAG is the foundation of every serious enterprise AI deployment.

2. Structured prompts. Instructions like "Cite your sources" or "Say I don't know if uncertain" reduce hallucination rates by 20% to 40% in benchmarks.

3. Reasoning models. Models that perform step-by-step reasoning before responding hallucinate less on complex tasks, though some reasoning models — notably OpenAI's o3 — show higher hallucination rates on person-specific questions.

4. Multi-model verification. Running the same query across multiple models and flagging disagreements catches errors no single model surfaces. Different models hallucinate different facts, so consensus signals reliability.

5. Human review for high-stakes content. No technique eliminates hallucination entirely. A 2025 mathematical proof showed that hallucinations are structurally inevitable under current LLM architectures.