Inference and generation

Hallucination

stable definition
Machine-readable Download Markdown

Definition

Hallucination is generated content that lacks support from the source or context that should govern the answer, conflicts with that source, or invents material and presents it as factual. Common examples include fabricated citations, nonexistent tool results, invented product behavior, and details added to a summary that do not appear in the source.

The boundary depends on the task. In source-grounded summarization, a true statement can still count as a hallucination if it came from outside the allowed source. In open-domain question answering, researchers may instead reserve the term for claims that conflict with established facts. Dark Factory Dev uses the broader operational definition because both failures require external verification.

History and disagreement

There is no reliable single coinage for the current NLP meaning. Simon Baker and Takeo Kanade's 1999 Carnegie Mellon report Hallucinating Faces is an early documented AI use of the word. It described learned pixels added during face super-resolution, a different meaning from unsupported text generation.

The term later spread through natural-language generation and then public discussion of chatbots. A 2024 peer-reviewed audit by Pranav Narayanan Venkit and coauthors found substantial disagreement across NLP research and practitioner definitions. Any metric or incident report that counts hallucinations should therefore state its evidence boundary and labeling rule.

Why fluency is a weak signal

Hallucinated content can be grammatical, specific, and internally consistent. Agent systems add another failure surface when a model turns a tool error into a plausible success narrative. The model's confidence, detail, or self-report does not establish that an action occurred. Logs, artifacts, source checks, tests, or other independent evidence must do that work.

Distinguish it from nearby terms

  • A factual error is any false statement. Whether every factual error is labeled a hallucination depends on the evaluation definition.
  • Nondeterminism means repeated runs may differ. A deterministic output can hallucinate, and a variable output can remain fully supported.
  • Deception implies an intent to mislead. Hallucination describes an output failure and does not establish intent.
  • Creative invention is expected when the task asks for fiction. It becomes a failure when the output presents invention as evidence or fact.

Check your understanding

An agent reports that a deployment succeeded and supplies a plausible URL, but no deployment log or reachable artifact exists. Identify the failure without using the fluency of the report as evidence either way.