AI Infrastructure

Hallucination

A hallucination is output from a language model that looks plausible and fluent but is factually incorrect, unsupported by source material, or fabricated entirely.

Hallucinations happen because language models are trained to produce likely-sounding text, not verified text. When the model is asked something it does not reliably know, it falls back on the patterns of similar statements it has seen before. The result reads as confident even when the underlying claim is invented. Classic examples include made-up citations, fake quotes, wrong dates, and fictional product features.

The failure mode shows up most often when the model is operating ungrounded, asked questions outside its training distribution, or fed contradictory or noisy source material. It is not a sign of a broken model. It is a sign that the model was asked to generate without adequate grounding.

The remedies are structural. Ground the model in retrieved source material (RAG). Track citations so every claim is verifiable (citation graph). Structure the source material into an ontology so retrieval is accurate. Prune the context window so the model is not reasoning over noise. Hallucination rates drop sharply when all four are in place and rise sharply when any one is missing.

The Amdahl view

Hallucination is a data problem pretending to be a model problem. Upgrading the model helps marginally. Grounding the model in a proper context layer with citations eliminates most of the failure mode. Teams still complaining about hallucination in 2026 have not invested in context engineering. The solution has been clear for two years. What is missing is the willingness to build the infrastructure.

See customer intelligence running on your own customer conversations.