Foundations

Hallucination

When an LLM generates plausible-sounding but factually incorrect output.

A hallucination is any confidently-stated LLM output that is not grounded in reality — a fabricated citation, a made-up API endpoint, a wrong date. It is the single biggest barrier to enterprise adoption of generative AI.

Hallucinations arise because LLMs are trained to produce fluent text, not to say 'I don't know'. Reducing them requires grounding (RAG), verification (tool calls, code execution), and calibration (letting the model express uncertainty).

Production systems reduce — not eliminate — hallucinations. The goal is to catch them before they reach users, through citations, confidence scores, and human review on high-stakes outputs.

Key points

  • RAG with strict citation requirements cuts hallucination rates 5–10x
  • Structured outputs with schema validation catch fabricated fields
  • Prompting the model to 'answer only from the provided context' helps but is not bulletproof
  • Frontier models hallucinate less but still hallucinate confidently

Common use cases

Legal, medical, and financial applications require the strictest grounding
Customer-facing chatbots benefit from citation UI
Internal tools can tolerate more risk with a 'verify before acting' UX

Frequently asked

Related terms

Applying Hallucination in your business?

We help enterprises design, ship, and govern AI systems end-to-end.