Hallucination rate is the percentage of an AI model's outputs that contain fabricated information presented as fact, details, sources, or claims that sound plausible but aren't actually true or don't exist. It's a way of quantifying how often a model confidently makes things up rather than admitting uncertainty or sticking to what it actually knows.

‍

‍Why it matters

Language models generate text by predicting likely word patterns, not by checking a database of verified facts, which means a fluent, confident-sounding answer and a factually correct one aren't the same thing and don't always come together. This becomes a serious problem in any application where someone might actually act on the output, a fabricated legal citation, a made-up statistic, or a nonexistent product feature can cause real harm precisely because the model states it with the same confidence as something true. Tracking hallucination rate is how teams put an actual number on a risk that would otherwise just be a vague concern.

‍

‍How it's measured

Measuring this typically involves having human reviewers fact-check a sample of model outputs against verified sources, flagging any claim that's fabricated or unverifiable, then calculating what percentage of outputs contained at least one hallucination. Because fact-checking every single output isn't practical at scale, teams often focus this kind of review on higher-risk categories first, like factual claims, citations, or specific numbers, rather than treating conversational filler with the same scrutiny.

‍

‍Where it's tracked

Customer-facing chatbots track this closely, since a hallucinated answer about a product or policy can directly mislead a user. Search and research tools built on language models track it too, especially when the tool cites sources, since a fabricated citation is arguably worse than no citation at all. Legal and medical AI tools face the highest stakes here, where a hallucinated case reference or drug interaction isn't just embarrassing, it's genuinely dangerous.