Hallucination
A hallucination is an AI-generated statement presented as factual even though the information available to the system does not support it.
How hallucinations happen
Language models generate likely continuations, not verified records. When a caller asks for a fact that is missing, ambiguous, or outside the agent's available context, the model may produce a specific-sounding answer instead of expressing uncertainty. A hallucination can be false, partly false, or accidentally correct; the defining problem is that the answer is not supported by the sources available for that response.
Several conditions make unsupported answers more likely. The prompt may reward answering every question. A knowledge search may retrieve an irrelevant passage. A business tool may fail or return incomplete data. Earlier details in a long conversation may be lost or confused. The caller may also ask a leading question that contains a false premise.
Why it matters for AI phone calls
Spoken answers are immediate and easy to act on. A caller may write down a fabricated appointment time, rely on an invented policy, or make a decision before anyone reviews the transcript. The natural rhythm of a voice conversation can also make a confident answer sound more authoritative than it is.
Reducing hallucinations requires more than telling an agent to be accurate. Useful controls include grounding responses in approved documents or live tool results, defining what the agent may and may not answer, validating tool inputs, and requiring confirmation before consequential actions. When evidence is missing, the safe behavior may be to say so, collect the caller's details, or hand the call to a person.
Testing should include questions that the source material cannot answer, conflicting information, tool failures, corrections from the caller, and attempts to push the agent outside its role. Reviewers should distinguish a retrieval failure from a generation failure: either can produce a bad answer, but they require different fixes.
No system can promise that hallucinations will never occur. The practical goal is to reduce their likelihood, limit what an unsupported response can change, and make failures visible enough to investigate.