Hallucination detection is the capability to identify when an AI system generates information that is factually incorrect, fabricated, or unsupported by its knowledge sources—before that information reaches the customer.

Because hallucinations are inherent to how language models work, detection becomes a critical safety layer. Effective detection approaches include: cross-referencing generated claims against source documents, confidence scoring that flags uncertain assertions, semantic consistency checks that catch contradictions, and specialized classifiers trained to identify hallucination patterns.

For customer service, hallucination detection must be real-time. Catching a hallucination after the customer receives it is too late—the trust damage is done. This requires inference-time detection integrated into the response generation pipeline, not post-hoc review.

Detection accuracy varies by domain. Factual claims ("your order ships tomorrow") are more verifiable than judgment calls ("this is the best option for you"). The most rigorous approaches flag anything that isn't directly grounded in retrieved sources, accepting some false positives to minimize false negatives on high-stakes information.

Related terms: AI hallucinations, AI grounding, Retrieval augmented generation