Should we call LLM errors hallucinations or fabrications?

The terminology we use for LLM errors shapes what fixes we build. The dominant terms are wrong in ways that matter.

Hallucination is a failure of perception — the experience of something as present in the world that is not actually present. LLMs do not perceive. They have no sensory access to the physical environment. Calling LLM errors "hallucinations" imports a perceptual model that doesn't apply and implies a fix (better perception/grounding) that addresses the wrong layer.

Confabulation is a psychological term. Human confabulation is the production of quasi-sensible narratives in response to queries that bear little relationship to the actual state of the world — a cognitive compensation for memory gaps, typically in neurological conditions. This is closer but still wrong. LLMs are not confabulating; they are not producing invented stories to fill gaps in an otherwise functioning memory. They have no memory that could gap.

Fabrication is the correct term. LLMs generate sensible-seeming text on the basis of processed corpus — statistical relationships between tokens. This text has no grounding in shared context or experience with the current interlocutor. The key point: accurate and inaccurate outputs are produced by the same process. There is no difference in the internal mechanism that generates a true statement versus a false one. Calling the false ones "errors" implies the true ones result from something different — they don't.

LLM text is fabrication even when the resulting output is appropriate and accurate to the reader's needs. This is not a judgment about the usefulness of LLMs but a description of the generative process. The absence of a grounding-in-shared-context is constitutive, not incidental.

The implication for safety: if we call errors "hallucinations," we build better perception. If we understand them as fabrication, we build better verification systems and calibrated uncertainty.

One concrete design response to out-of-domain fabrication is the explicit domain rejection pattern: when a query falls outside the defined domain scope, the system generates a structured "this topic lies outside my designed domain" response rather than attempting to fabricate a domain-appropriate answer. The FALM system (business and media domain LLM) implements this as a hard boundary — the model has a defined domain scope and generates rejection responses for out-of-scope queries rather than producing ungrounded outputs. This is the design intervention that treats fabrication as a fixed property rather than a failure to fix: instead of trying to stop the model from fabricating, you define the boundary within which fabrication is acceptable (because you can verify the training distribution) and explicitly block out-of-domain queries. See also Why do specialized models fail outside their domain? — the rejection pattern is one engineering response to the cliff problem.

Fine-grained fabrication taxonomy (FAVA): Not all fabrication is the same. FAVA decomposes fabrication into six types requiring different verification strategies: (1) entity errors — incorrect single-entity references, fixable by single-source lookup; (2) relation errors — wrong relationships between correctly identified entities; (3) sentence-level contradictions — claims that conflict with source material; (4) invented entities — references to non-existent things, requiring multi-source verification because no single source confirms a negative; (5) subjective claims presented as facts; (6) unverifiable statements. The practical implication: a single "hallucination detector" is the wrong architecture. Each fabrication type has different verification costs and strategies. Entity errors are cheap to catch; invented entities are expensive; subjective claims require a different evaluation framework entirely.

Falsification beyond fabrication: The Knowledge Custodians analysis adds a dimension to the fabrication taxonomy that the FAVA categories don't capture: rhetorical falsification. AI doesn't only fabricate (produce outputs without grounding) — it also uses unintended linguistic sleights of hand. These techniques of persuasion, cognitive techniques and discursive techniques, falsify and mislead without making up non-facts. They use rhetorical devices, commonly-used manners of speech and argumentation, patterns of confident expression learned from SFT examples and RL, to generate responses that appear more professional and convincing than they are in substance. This is a seventh category beyond FAVA's six: claims that are technically true or at least not fabricated, but whose rhetorical framing implies a certainty, authority, or insight that the generative process did not produce. Since Does RLHF training make models more convincing or more correct?, the alignment process systematically amplifies this category. The model learns the form of confident expert expression — hedging patterns, qualification structures, authoritative framing — and applies these forms to content that doesn't warrant them. The output reads as expert judgment, but the judgment was never performed.

Source: Linguistics, NLP, NLU; enriched from Domain Specialization; enriched from Self Refinement Self Consistency Feedback; enriched from inbox/Knowledge Custodians.md