Hallucination

Hallucination is the failure mode where a language model generates content that is fluent and confident but factually wrong. The model invents a citation that does not exist, states an incorrect date, misattributes a quote, or confabulates technical details. The output is produced by the same next-token-prediction mechanism that produces correct answers; from the model's perspective there is no distinction between generating a known fact and generating a plausible-sounding fabrication. This makes hallucination the central reliability concern for production generative AI and the reason most enterprise deployments pair LLMs with retrieval, verification, or tool use rather than relying on pure pretrained recall.

The taxonomy matters because different kinds of hallucination call for different defences. **Factual hallucination** — wrong dates, names, numbers — is addressed by grounding responses in retrieved documents (RAG) or authoritative tools (search, calculators, databases). **Reasoning hallucination** — plausible-looking chains of thought that reach incorrect conclusions — is addressed by self-consistency sampling, verification models, or tool use that forces the model to actually compute rather than estimate. **Instruction hallucination** — claiming to have done something the model cannot ("I have sent the email", "I have checked the database") — is addressed by explicit tool-use architectures so the action either happens or does not. **Citation hallucination** — fabricated references to papers or sources — is addressed by retrieval with quoted passages and URL verification.

For APAC mid-market enterprises, the practical posture is to assume unmitigated LLM output will hallucinate on factual questions and design systems accordingly. **Retrieval grounding** handles long-tail factual knowledge. **Structured tool use** handles anything involving live data or action. **Output verification** (schema validation, citation checking, value-range checks) catches residual hallucinations before they reach users. **User-facing transparency** — showing sources, flagging low-confidence answers, distinguishing summary from verbatim quotes — lets users calibrate their own trust in model output.

The non-obvious operational note: **hallucination rates vary by topic in predictable ways**. Models hallucinate far more on rare, recent, or non-English-dominant topics than on common English-language subjects well-represented in pretraining. If your workload is Japanese legal terminology, Korean regulatory policy, or Indonesian tax code, the baseline hallucination rate will be materially higher than English benchmarks suggest. Measure hallucination on your actual workload distribution, not generic benchmarks, and invest in retrieval for the long-tail, non-English-heavy topics specifically.