AI / Voice
Hallucination
An AI hallucination, also called a confabulation, is a plausible-sounding response that is false, unsupported, or inconsistent with the available source information.
An AI hallucination, also called a confabulation, is a plausible-sounding response that is false, unsupported, or inconsistent with the available source information.
An AI hallucination is a response that sounds plausible but is false, unsupported, or inconsistent with the source material. NIST uses the term confabulation for this behavior and notes that generated content may confidently present incorrect information. In a voice workflowWorkflowA workflow is a defined sequence of steps, decisions, actions, delays, and outcomes used to complete a business process., the risk is practical: a caller may act on a wrong policy, price, appointment detail, or eligibilityEligibilityEligibility is the set of conditions that determines whether a prospect can move to a service, quote, appointment, application, or specialist. answer.
Why hallucinations happen
A language model generates likely text from patterns and context. It is not a database that automatically knows which statement is current or approved. Vague prompts, missing knowledge, conflicting sources, overly broad questions, and requests outside the agent's job all increase the chance of an unsupported answer. The NIST Generative AI Profile treats confabulation as a risk that must be managed, not a behavior that can be assumed away.
How teams reduce the risk
- Ground answers in a maintained knowledge base or other approved system of record.
- Use exact messages and deterministic rules for critical statements and decisions.
- Define topics the agent must refuse or escalate.
- Test realistic caller phrasing, edge cases, and conflicting information before launch.
Thoughtly's Genius knowledge baseKnowledge baseA knowledge base is a maintained collection of approved business information that an agent can use to answer questions. can supply retrieved business information during a conversation, while rules and handoffs control what happens when an answer should not be generated. These controls reduce risk, but they do not make every output automatically correct.
Frequently asked questions
Is a hallucination the same as a transcription error?
No. A transcriptionSpeech-to-Text (STT)Speech-to-Text (STT) converts spoken audio into text that a voice agent can interpret, store as a transcript, and use in downstream actions. error misrepresents what the caller said. A hallucination is generated content that is unsupported or false. One can contribute to the other.
Can RAG eliminate hallucinations?
No. Retrieval improves grounding, but outdated, irrelevant, or contradictory source content can still produce a wrong answer.