Overview
When voice agents make up false information, invent policies, or provide incorrect facts during calls.
Why it matters
Critical reliability issue: voice agents inventing return policies, quoting wrong prices, or making promises your business can't keep.
How it works
Hallucination works by processing voice data through multiple stages of the AI pipeline, from recognition through understanding to response generation.
Common issues
Hamming's guardrail testing specifically catches hallucinations where agents provide false information confidently. Their LLM-based evaluations verify factual accuracy and flag responses that deviate from ground truth.
Implementation guide
Test for hallucinations using Hamming's approach: create test cases with known correct answers, validate agent responses against ground truth, monitor hallucination rates in production, and implement guardrails to prevent common hallucination patterns.
Hamming's benchmarks
Drawing on Hamming's 10M+ mins protected across 10K+ voice agents:
| Metric | Excellent | Good | Acceptable |
|---|---|---|---|
| Hallucination Rate | <5% | Not set | Not set |