Evaluation

Hallucination

When a model states something false with full confidence — the failure mode grounding, citations, and evals exist to contain.

Operating principle

Production AI is a system: context, tools, permissions, traces, evals, and feedback loops around the model.

What it is

A hallucination is a fluent, confident, wrong answer: an invented policy, a citation to a document that doesn't exist, a number that was never in the source. It happens because language models generate plausible text; plausibility and truth usually overlap, and a hallucination is where they don't.

Why it matters

For enterprise AI this is the failure mode that costs trust fastest. One invented refund policy told to a customer outweighs a hundred correct answers. The goal in production is not a model that never errs — none exists — but a system where wrong answers are rare, detectable, and cheap to correct.

How we contain it

Grounding: answers draw from retrieved documents, not the model's memory. Citations: every claim links to its source, and unverifiable citations are flagged rather than shown. Refusal: the assistant says 'I don't know' when retrieval comes back thin. Evals: groundedness is scored continuously, so regressions surface as numbers instead of incidents.

Related resources