What causes AI hallucinations?

What causes AI hallucinations?

An AI hallucination is generated content that is false, unsupported, or inconsistent with the relevant evidence. Examples include an invented book citation, an incorrect summary detail, or a function name that a software library does not provide.

The term describes a problem in the output. It does not require assuming that the system has human-like intentions or experiences.

Plausible continuation is not verification

Language models learn to produce sequences that fit patterns in their training and context. A sequence can sound natural without corresponding to a true statement.

For an obscure question, the model may have insufficient information but still generate a familiar-looking answer structure. A convincing citation format, for example, does not establish that the cited publication exists.

Missing and misleading information matter

An ambiguous question can lead the model toward the wrong interpretation. Incomplete context can leave it to fill gaps, while incorrect source material can contaminate an otherwise coherent answer.

Retrieval can help by supplying evidence, but the retrieved material may be irrelevant or misunderstood. OpenAI's accuracy guidance distinguishes failures in retrieval from failures in how a model uses available context.

Diagnose the specific failure

Suppose a generated summary says a meeting approved a launch date when the transcript only discusses a proposed date. The error is not solved by making the prose more polished. The system needs to preserve the difference between a proposal and a decision.

Useful checks include inspecting sources, testing calculations, and identifying claims unsupported by the supplied material. Asking the model to acknowledge missing information can help, but evaluation is still needed to see whether that behavior is reliable.