What Is an AI Hallucination?

A hallucination is output a language model presents as fact but that is fabricated or unsupported. It happens because a model predicts plausible continuations of text, not verified truth — fluency and accuracy are separate properties, and nothing in the base objective rewards the model for abstaining when it does not know. For an autonomous agent acting on its own output, an unflagged hallucination is not a quirk but an operational risk.

What it is

Fabricated output presented as fact

Root cause

Models predict plausibility, not truth

Mitigation

Retrieval, citation, and verification

Why fluency is not accuracy

A model is trained to produce text that looks right, and most of the time looking right and being right coincide. When they diverge — an obscure fact, a fabricated citation, a confident wrong number — the model has no built-in signal that it has crossed the line, so it states the fabrication as fluently as the truth.

The estate’s answer is structural rather than hopeful: ground answers in retrieved, citable sources, and keep a human in the loop for consequential actions, so an agent’s confidence is never the only thing standing behind a decision.

Related standards

Survey of Hallucination in NLG — Ji et al., 2022

Questions

Can hallucinations be eliminated?

Not entirely with current models. They are reduced by retrieval grounding, citation, verification steps, and bounding what an agent can act on unaided.

Is a hallucination a bug?

It is a property of how the models work, not a defect to patch — which is why systems are designed to contain it rather than assume it is gone.