In short: Your AI hallucination rate is the share of answers that contain something false or made up, measured on a fixed set of questions. A model’s score on a public leaderboard isn’t your rate, because your prompt, documents and users are different. Measure it yourself with real questions, run each one several times and watch the trend after every change.
You put an AI assistant to work so customers and colleagues get quick, correct answers. Most of the time it does just that. Now and then it says something that sounds right and isn’t. You’d like to know how often. The AI hallucination rate gives you that number, if you measure it in a way that fits your own setup. Here’s how.
What is an AI hallucination rate?
An AI hallucination rate is the percentage of answers in which the AI states something false or unsupported as if it were true. You measure it on a fixed set of test questions. If 6 of 100 answers contain a made-up fact, your rate on that set is 6%. That number only means something together with the questions behind it. Easy questions give a low rate and hard ones a high rate, so always keep the test set next to the score. Researchers use different words for the same problem. The US standards institute NIST calls it confabulation in its profile of generative AI risks. OWASP lists it under misinformation in its top ten risks for apps built on language models. For the basics of why it happens, read our post on AI hallucinations.
Why isn’t the leaderboard rate your rate?
Public leaderboards measure a model on one task that someone else chose. Vectara’s well-known hallucination leaderboard, for example, asks models to summarise short documents and checks whether the summary adds facts that aren’t in the text. That’s useful when you choose a model. It doesn’t tell you how your own assistant behaves.
Your assistant works with your system prompt, your documents and your customers’ questions. A model that scores well on summaries can still invent a refund rule that isn’t in your help pages. It can also mix up two of your products or guess a price. The other way round, a modest model with good documents and tight instructions can do better than the leaderboard suggests. The only way to know your rate is to measure it on your own questions.
How do you measure your own hallucination rate?
You measure it with a fixed test set that you run again and again. It takes a day to set up and is easy to repeat after that.
- Collect real questions. Take 50 to 100 questions your users actually ask, with answers you know are correct.
- Add questions with no answer. Ask about a product you don’t sell or a rule that doesn’t exist. The right answer is “I don’t know” or “we don’t offer that”.
- Add tricky cases. Questions near the edge of your documents, questions with a false premise, and questions that mix two topics.
- Score every answer in one of four groups. Correct. Wrong. Made up, meaning it invents a fact, a source or a rule. Honest refusal, meaning it says it doesn’t know.
- Run every question several times. A language model doesn’t always give the same answer, so one run can mislead you.
Your hallucination rate is the share of answers in the “wrong” and “made up” groups. Track the made-up group on its own as well, because that’s the one that hurts trust most. Have someone who knows the subject check the scores. An AI can help sort the answers, but a person should decide the close calls.
What rate is good enough?
There’s no single good number. It depends on what a wrong answer costs. A wrong opening time is annoying. A wrong answer about a contract, a medicine or a refund can cost real money or harm someone. So set a limit per topic. Be strict where answers carry weight, and allow a bit more where they don’t.
Two things matter more than the number itself. The first is confident wrong answers on serious topics. Those should be close to zero. The second is the trend. Even well-designed AI systems drift and misbehave over time. A new model version, a changed prompt or a new document can push your rate up without anyone noticing. Our post on AI monitoring explains how to keep an eye on that once your AI is live.
There’s a trade-off too. If you push the rate towards zero, the assistant tends to refuse more and help less. Watch the honest refusals next to the hallucinations, so you see both sides.
What this means for your team
Start small. Build a set of 50 questions this week, including ten with no answer. Run it five times and write down your starting rate. Then run the same set after every change and at least once a month. Keep the answers, not just the score. They show your team or your vendor exactly what to fix.
Doing this by hand gets heavy once you want to test every week. VeritasAI runs more than 90 automated risk probes covering hallucinations, jailbreaks and data leakage. The PRO plan runs them on a weekly schedule. Every run keeps the exact responses your model gave, so you can check each score yourself.
Know your hallucination rate every week, not just once.
Frequently asked questions
Which AI model hallucinates the least?
It changes with every new release, and it depends on the task. Leaderboards give a rough picture for picking a model. Your own test set tells you how it performs on your questions.
Can you get the hallucination rate to zero?
Not with today’s language models. You can get it low on the topics that matter by giving the AI good documents, clear limits and permission to say it doesn’t know. Then keep measuring.
Does RAG stop hallucinations?
RAG means the AI looks up your documents before it answers. It helps a lot, but it doesn’t stop hallucinations. The AI can still misread a document or fill gaps with a guess.
How many test questions do you need?
Fifty to a hundred is a good start for one assistant. Add questions whenever a real mistake turns up, so the set grows with what actually goes wrong.