Risks & safety
Hallucination
Also known as: confabulation
When an AI produces information that sounds confident and plausible but is actually false or made up.
A hallucination is when an AI confidently tells you something that simply is not true, a fake fact, a made-up quote, a book that does not exist. It is not lying on purpose; the AI has no idea what is true. It writes what sounds right based on patterns, and sometimes that plausible-sounding text is just wrong. This is the single most important reason to check anything important an AI tells you.
A hallucination is a confident but false or unsupported output from a generative AI system. It happens because a language model is trained to produce likely-sounding text, not verified truth, so a fluent, well-formed sentence can still be factually wrong.
Common forms
- Invented facts, dates or statistics.
- Fake citations, quotes or references.
- Plausible but incorrect answers stated with full confidence.
How to reduce it
- Ground the model: use retrieval-augmented generation so answers come from real documents.
- Prompt carefully: ask for sources, allow "I don't know", and give the needed context (prompt engineering).
- Verify: always check facts, figures and citations before relying on them.
Hallucination denotes generated content that is unfaithful to the source or to world facts despite being fluent and often confident. A common taxonomy distinguishes intrinsic hallucination (contradicting provided input) from extrinsic (unverifiable against input) (Ji et al. 2023). The root cause is structural: likelihood-maximising decoding over a parametric LLM optimises plausibility, not veracity, and the model has no built-in truth signal or calibrated notion of its own uncertainty.
Mitigations operate at several layers: grounding via retrieval and tool use, decoding and self-consistency checks, alignment training that rewards calibrated abstention, and post-hoc verification or citation. None eliminate the failure mode, which is why human oversight remains essential for high-stakes use, and why evaluation increasingly measures faithfulness and attribution rather than fluency alone. Note the term is contested: some researchers prefer "confabulation" to avoid implying perception.