Hallucination Detection

LLMs hallucinate. This is not a bug to be fixed in the next model release — it is an inherent property of probabilistic text generation that the industry is actively working to reduce but has not eliminated. Enterprise AI deployments must treat hallucination as a known failure mode and build detection and mitigation into the production architecture.

Hallucinations fall into two categories: **intrinsic** (the model contradicts information present in its own context or source documents) and **extrinsic** (the model generates claims that cannot be verified against any provided source). Detection approaches range from simple — checking whether the model's response cites passages that actually appear in the retrieved context — to sophisticated: using a second LLM as an evaluator to score faithfulness, using natural language inference (NLI) models to assess entailment between response claims and source documents, or fine-tuning a dedicated hallucination classifier on labeled examples.

The enterprise response to hallucination risk has three layers: **prevention** (RAG architectures that ground responses in retrieved evidence, conservative system prompts, and temperature calibration); **detection** (automated faithfulness scoring on every production output using LLM-as-judge or NLI models); and **mitigation** (flagging low-confidence responses for human review, refusing to answer when evidence is insufficient, and citing sources to enable user verification). Organizations in legal, medical, and financial sectors should layer all three.