Hallucination

Hallucination is a failure mode of large language models (LLMs) where the model generates plausible-sounding but factually incorrect or entirely fabricated information — often with high confidence, as if it were true.

Why Hallucinations Happen

LLMs are pattern-matching systems trained to predict the next word in a sequence. They have no inherent mechanism to verify truth. If training data contains misinformation, or if a question falls outside the training distribution, the model may "fill in gaps" by inventing details rather than saying "I don't know."

Examples in Voice Agents

• Caller asks about a new product; AI invents features that don't exist.
• Caller asks their account balance; AI generates a plausible-sounding number instead of querying the database.
• AI incorrectly tells a patient their medication is safe for their specific condition.

Mitigating Hallucinations

• Retrieval-Augmented Generation (RAG): Ground AI responses in verified documents and database queries.
• Confidence thresholds: If AI is uncertain, escalate or say "I don't have that information."
• Structured outputs: Force AI to query systems (CRM, EHR) rather than generating answers from memory.
• Human review loops: Critical information (medical advice, account numbers) should be verified by a human before delivery.

Workforce Wave Hallucination Prevention

Workforce Wave uses RAG and structured integrations with CRM/EHR systems to ground voice agent responses in verified data, not model memory. This eliminates hallucinations for account lookups, appointment booking, and health information delivery.