Glossary

Hallucination

In generative AI, hallucination refers to model outputs that are presented as coherent content yet are unsupported by, fabricated relative to, or inconsistent with the relevant source evidence or task constraints — for example invented citations, false facts, or unfaithful summaries. It is a reliability failure mode of generative systems; it is not proof of model “intent,” and it is distinct from ordinary uncertainty messaging, mere staleness, systematic bias, and malicious prompt injection.


Why it matters

Enterprise workflows that treat fluent text as verified knowledge amplify hallucination into operational and legal risk. Mitigation patterns include retrieval grounding, citation checks, constrained decoding, human review for high-impact acts, and evaluation suites that score faithfulness — not only eloquence.

Distinctions: outdated training knowledge can produce wrong answers that are not “fabricated” in the same sense as invented sources; bias can be systematically skewed yet grounded in training patterns; prompt injection is an attacker-driven control failure. Anthropomorphic language (“the model lied”) obscures the engineering problem: sampling fluent strings without guaranteed grounding.

Some technical literature uses “confabulation” for related fluent-but-ungrounded generation. In this glossary, hallucination is the canonical operator term; confabulation is related vocabulary, not an exact alias.

Hypothetical illustration: An analyst asks an LLM for case-law support and receives plausible-looking citations that do not exist. A RAG-backed assistant that must quote from an approved corpus reduces that failure mode but can still miscompose retrieved snippets — so review remains required before filing.

Mechanisms

  • Fluency without grounding — decoding optimizes for likely text, not guaranteed correspondence to evidence.
  • Context and objective mismatch — prompts that demand specifics beyond available evidence invite fabrication.
  • Mitigation stack — RAG, tool-verified claims, constrained outputs, and human-in-the-loop checks for consequential uses.

Boundaries

  • Hallucination ≠ every incorrect answer (include calibration, stale knowledge, and user misunderstanding).
  • Hallucination ≠ model bias as exact synonyms.
  • Hallucination ≠ prompt injection (security/control problem) though attacks can induce unreliable outputs.
  • Hallucination ≠ confabulation as exact synonyms in TSS Knowledge — confabulation appears in some technical literature as related language for ungrounded fluent output; this entry canonicalizes hallucination.
  • Avoid claims that the model “knows,” “believes,” or “intends” when describing hallucination.

Sources

  1. Ji, Ziwei, et al. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys — overview of hallucination phenomena and evaluation in NLG.
  2. Maynez, Joshua, et al. (2020). On Faithfulness and Factuality in Abstractive Summarization — https://arxiv.org/abs/2005.00661 — faithfulness/factuality failures related to hallucinated content.
  3. Lewis, Patrick, et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — https://arxiv.org/abs/2005.11401 — grounding pattern commonly used to reduce unsupported generation (not a complete cure).



Further research


Continue exploring

A short path from this concept into related research and practical resources.

  1. Related conceptGenerative AI
  2. Deeper researchWhy Ai Features Dont Compound
  3. Apply in DiscoverThe Enterprise AI Playbook Library
AI Hallucination — Definition | The Strategy Stack