What Is an AI Hallucination?

Hallucinations take several forms: fabricated details, invented citations, made-up policies, or a confident answer to a question the model has no information about. The underlying cause is structural, since a model trained to produce likely text will produce likely-sounding text even when it lacks the knowledge to be correct. Fluency is not evidence of accuracy.

The most effective mitigation is grounding: retrieve real source passages and instruct the model to answer only from them, then instruct it to say it does not know when the sources do not cover the question. Escalating uncertain cases to a person and giving operators a way to correct wrong answers close the remaining gap.

For a business-facing chat widget the stakes are practical, because an invented price or policy becomes a customer complaint. Clerkzo constrains answers to content crawled from the customer's own site, routes anything outside that to a human, and gives owners a corrections loop so a wrong answer can be fixed permanently.