What counts as a hallucination
Two kinds of failure share the name. The first is fabrication: the model asserts something about the world that is not true — a court case that never happened, a paper with real-sounding authors that was never written, a product feature that does not exist, a statistic with a precise decimal and no origin. The second is unfaithfulness: the model was handed a document and says something the document does not say, or contradicts it. The second kind is especially costly in business use, because the whole point of supplying the document was to keep the answer tied to it.
What both have in common is tone. Models do not hedge more when they are wrong; an invented answer arrives in the same fluent, assured register as a correct one. That is what separates a hallucination from an ordinary mistake a reader would notice, and why it cannot be caught by reading the answer alone.
Why language models hallucinate
A language model is trained to continue text plausibly, not to check facts. For things that appear many times in its training data, the plausible continuation and the true one coincide. For rare facts — a minor person's birthday, a niche figure, a specific page number — the model has nothing reliable to draw on, and it still produces the most plausible-looking answer.
A September 2025 paper by Adam Tauman Kalai and three co-authors argues that this is not mysterious at all. In their account, hallucinations are ordinary prediction errors that arise naturally during pretraining whenever a model cannot tell a true statement from a false one. They persist after training for a reason outside the model: most widely used evaluations score an abstention the same as a wrong answer — zero — so a model that guesses scores better than one that admits uncertainty, much like a student guessing on a multiple-choice exam. The authors' proposed fix is to change how mainstream benchmarks score uncertain answers, so that honesty stops being penalised.
How to reduce hallucinations
Anthropic's guidance for Claude is a good summary of what works across models:
- Allow “I don't know”. Explicitly permitting the model to say it lacks the information removes some of the pressure to invent.
- Ground the answer in supplied text. Give the model the source documents and tell it to use only those. For long documents, Anthropic suggests having the model first pull out word-for-word quotes, then answer from the quotes.
- Require citations, then check them. Ask for a supporting quote for each claim, and have the model withdraw any claim it cannot support.
- Compare several answers. Run the same prompt more than once; where the answers disagree, something is likely made up.
- Inspect the reasoning. With extended thinking turned on, the reasoning trace can show where a wrong assumption entered.
Retrieval systems (RAG) and search tools apply the grounding idea at scale. None of these remove hallucinations entirely — Anthropic says so directly — so anything high-stakes still needs a human check.
What reducing hallucinations costs
Every defence on that list has a token price, and it is worth knowing which part of the bill each one moves:
- Grounding adds input: the source documents travel with every request. Retrieval keeps this down by sending only the relevant passages, and prompt caching discounts a document reused across many questions.
- Quotes and citations add output — the most expensive token type on most models.
- Comparing several answers multiplies the whole call: three samples cost three times one.
- Reasoning adds reasoning tokens, billed as output even though you never see most of them.
- Web search adds per-search fees on top of tokens at some vendors — Anthropic, for example, charges per 1,000 searches — plus the retrieved pages as input on later turns.
The worked example prices one question at three levels of checking on the same model. The point is not that checking is expensive; it is that it should be sized to the stakes. A support bot answering from a product manual needs grounding; a legal brief needs grounding, citations and a lawyer.
Worked exampleOne question, three levels of checking
10,000 requests a month on Claude Sonnet 5.5, priced at its current list rates, checked on 2026-10-09.
10,000 questions a month. All token counts are illustrative assumptions, and reasoning-token volumes in particular vary widely by task and model. Grounding is billed as input; quoted citations and reasoning are billed as output.
Change the workload or the model in the LLM API cost calculator.
Where hallucinations do the most damage
The damage depends less on how often a model hallucinates than on what happens to the output next. Text a human reads critically is low risk. Output that is copied into a legal filing, a medical note, a financial report or a customer reply without review is high risk, because the error leaves the system with your name on it. Code is a special case: an invented function or library name usually fails loudly at run time, but a plausible-looking wrong calculation does not. Design the review step around where the output goes, not around the model's reputation.
Measuring hallucination in your own system
Public leaderboards measure models on someone else's questions. What matters is the rate on yours. Build a test set of realistic questions with known answers, including some the system should decline because the information is not available. Score three outcomes separately — correct, wrong and abstained — rather than collapsing them into accuracy, so that a change which makes the system guess more is visible as more wrong answers rather than hidden as fewer abstentions. That is the same scoring principle Kalai and colleagues recommend for public benchmarks.