Hallucination
Hallucination is when a language model produces information that sounds plausible but is wrong or unsupported; for voice agents it is a risk to plan for.
Hallucination is when a large language model produces information that sounds plausible and confident but is wrong or not grounded in any source. When the model does not know an answer, it may build a fluent reply instead of saying "I don't know". In a written chat, the reader can spot and check this; on the phone, callers tend to trust what they hear. That is why hallucination is a risk to deal with before a voice agent goes live.
How it works
Language models generate the most likely next words in a text; they have no separate mechanism for checking whether what they produce is true. If the answer is not in the context the model was given, it may assemble one from general knowledge or similar examples. Typical cases in a voice agent are confirming an appointment slot that does not exist, quoting information that is out of date, or describing a service the business does not offer.
Why it matters
Wrong information given over the phone can turn straight into a complaint or a lost customer. The main safeguards are: grounding the agent in an approved knowledge base; writing an explicit "if you are not sure, don't guess, hand over" instruction; having details such as appointment availability looked up in a connected system rather than recalled by the model; asking the caller to confirm before any action; testing with awkward questions before go-live; and reviewing call transcripts regularly.
In Speakline
A Speakline voice agent answers from your knowledge base. In the agent's instructions you define that, when it is unsure, it logs the request or hands the call to a person with a summary of the conversation (a warm transfer). Calls are stored with their transcripts, so you can review exactly what the agent said.