One assistant cites a court ruling that does not exist. Another invents an annual report, complete with page number and figure. The tone is the same as for a correct answer. That is what we call a hallucination: a false statement, presented as true.
Why the model makes things up
An LLM produces the most plausible continuation of a text. When it knows the answer, the plausible continuation is usually the right one. When it does not, the plausible continuation still has the shape of an answer: a name, a date, a reference.
On an example sentence, the model picks one of the likely tokens, adds it to the sentence, then starts again: the most plausible continuation, not the most accurate one.
Key concept
A model does not volunteer that it does not know: its training and its evaluations reward an answer over an admission of ignorance. So it often states a guess as confidently as a fact. You can reduce hallucinations; you cannot remove them.
With exam grading, the model that guesses scores 4 points and the one that says “I don't know” 3 points: guessing pays.
The risk rises when the question is about a precise, rare fact: a figure, a quote, a source, a recent event. It falls when the model works from a document you provide rather than from memory. Anthropic's documentation details techniques to reduce hallucinations, and states that none of them eliminates them.
From memory: no passage is retrieved, the model still answers just as confidently, and there is nothing to check it against.
What it means for you
For founders
A hallucination in front of a customer costs more than the time it saved. Find the answers in your product where a mistake has a real cost, and check those, not everything. In a regulated sector, that list grows fast.
For product managers
Design the interface with the error in mind. Show the sources, let users open them, and give them a simple way to flag a wrong answer. When the assistant replies "I couldn't find it", that should be planned, tested behavior, not a failure.
For business leaders
Decide who checks what before AI-generated content leaves the company. An internal memo reread by its author does not carry the same risk as a customer letter or a figure in an annual report. Across a group, write that rule once, and make sure every team knows it.
Go deeperWhy guessing gets rewarded
A September 2025 research paper, "Why Language Models Hallucinate", offers a simple explanation. During training, a model learns to produce plausible text, not to separate true from false. Then most evaluations grade answers like an exam: a right answer scores points, an "I don't know" scores none. A model that guesses therefore scores better than one that admits it does not know. The authors propose changing those grading schemes rather than adding new tests. They also point out that a freshly pretrained model is often well calibrated (its confidence matches its rate of right answers), and that the training that follows can degrade that calibration.
The classic mistake
fAIling RULE VII
The trap: Hallucinations are just creative suggestions
Treating an invented source as a "creative suggestion" is the surest way to publish it. A quote, a figure or a reference produced by a model gets checked before it goes out, especially when it looks perfect.
Read rule VII in the manifestoTry it
Ask an AI assistant for five precise references on a topic you know well: articles, reports or rulings, with author and date. Check each one. Then ask again, adding "If you are not sure a reference exists, say so", and compare the two lists.
Around this topic
Go deeperGrounding answers and citing sources
The most reliable way to cut down on invention is to give the model the documents to use, and ask it to stick to them. Some APIs make this checkable: as of October 2026, Claude's citations feature returns the exact passages its claims rest on. Users can then check the answer instead of taking it on trust. That does not guarantee the model understood the passage, but it makes the error visible.
Go deeperThe cut-off date
A model knows nothing about what happened after its cut-off date, the limit of its knowledge. Asked about a newer price, law or product, it may answer with what used to be true, or invent a plausible continuation. The closer a question is to current events, the more you need to supply up-to-date information. The large language model (LLM) explains the difference between training and use.
Go deeperLetting the model say “I don’t know”
By default, a model answers. As of October 2026, according to Anthropic's documentation, explicitly allowing it, in its instructions, to admit uncertainty sharply reduces false information. The product then needs to know what to do with that "I don't know": hand over to a human, open a help page, or run a search.
Go deeperWeb search tools
A web search tool lets the model fetch recent information instead of guessing it. As of October 2026, with Claude's web search tool, the answer cites the pages it used. But a web page can itself contain errors, and the model can summarize it badly. Search moves the checking work; it does not remove it.
What to do with it
List the answers in your product where a mistake has a real cost, and check those rather than everything.
Show the sources, let users open them, and give them a simple way to flag a wrong answer.
Make "I couldn't find it" a planned, tested behavior, and decide who checks what before AI-generated content leaves the company.