There is no single, reliable frequency figure for how often AI chatbots make things up, and anyone quoting you a precise percentage is almost certainly pulling it from a specific test that does not generalize to your use case.
The honest answer is that hallucination rates vary enormously depending on the model, the task, the prompt, and whether the tool has access to real documents. What matters more than a headline number is understanding why the rate swings so widely, so you can predict when you are in a high-risk situation and build a habit that catches the errors before they cost you something.
Start with the mechanism. A large language model generates text by predicting what word or phrase most plausibly comes next, based on patterns learned during training. It has no built-in fact-checker and no internal flag that says "I don't actually know this."
When you ask about something well-represented in its training data — say, the capital of France or the syntax of a Python for-loop — the plausible next token is also the correct one, so it looks reliable. When you ask about something rare, recent, or highly specific — a niche product's return policy, a statistical figure from a small study, a citation — the model still produces fluent, confident-sounding text, but the probability of it being right drops sharply.
That is why the same chatbot can nail a complex coding question and then invent a book that does not exist in the same conversation.
Here is a concrete example of how this plays out in practice. Suppose you ask a chatbot: "What was the exact wording of the refund clause in the terms of service for [a small SaaS tool]?" The model has likely never seen that document.
It will often generate a plausible-sounding clause with specific-sounding language, because that is what a terms-of-service refund clause typically looks like. If you then ask the same model to summarize a document you paste directly into the chat, the error rate collapses, because now the answer is grounded in text the model can actually see.
Same tool, same session, wildly different reliability. That gap — between closed-book questions and grounded questions — is the single most useful thing to understand about hallucination.
So what should you actually do? First, treat any specific factual claim from a chatbot as a draft, not a source — especially numbers, dates, quotes, legal or medical details, and citations. Second, when accuracy matters, switch from closed-book to grounded mode: paste the source document, use a tool with web search enabled, or use a retrieval feature that pulls from a defined set of documents.
Third, ask the model to quote the exact sentence it is basing a claim on. If it cannot produce a verbatim quote from a source you provided, treat the claim as unverified. A useful tip that goes beyond the obvious: ask the same question twice in separate chats and compare the answers.
Fabricated details tend to drift between runs — the invented citation changes, the made-up statistic shifts — while genuinely known facts stay stable. That instability is a cheap, practical signal.
Now the limits, stated plainly. No reliable frequency figure exists in the sources available to us for how often any given chatbot hallucinates, and that is not a gap we can paper over with a made-up number. According to our AI tool database, which maintains pricing and capability snapshots for 360 AI tools with a most recent verification date of 2026-09-18, the snapshot records what each tool costs and what it can do — it does not record accuracy rates or hallucination frequency, and no vendor-published reliability metric is included.
That means the honest position is: you cannot look up a trustworthy failure rate, and you should not trust anyone who claims one without showing you the exact test conditions. The practical consequence is that your verification habits have to do the work that a reliability score would otherwise do.
This advice also breaks down in a specific way: grounding reduces hallucination but does not eliminate it. A model summarizing a document can still misread a table, merge two sections, or drop a qualifier. For anything with legal, medical, or financial consequences, a human who knows the domain still has to check the output.