A cat can stare at an empty corner with extraordinary conviction. A chatbot can do something similar, except it supplies three paragraphs and a bibliography. The unnerving part is how much the result resembles a sensible answer.
AI hallucinations are false or unsupported outputs presented in the course of generating an answer. They can be polished, specific and persuasive. Understanding them does not mean refusing to use AI; it means knowing which parts of its work need something stronger than an impressive writing voice.
What counts as an AI hallucination?
There is no single taxonomy used everywhere. NIST’s generative AI risk profile uses “confabulation” for erroneous content and includes answers that depart from their input or contradict earlier statements. Common examples include a nonexistent citation, a wrong date, or a summary that adds a claim absent from its source.
The problem is not necessarily that the model has copied a false statement from somewhere. It can assemble a new falsehood. Nor is the definition simply “something missing from the training data”: a correct inference can be new, while a well-represented myth can still be false.
Confidence makes mistakes harder to notice, but a hedge does not turn a false claim into a correct one. “Probably” followed by an imaginary journal article still leaves you with an imaginary journal article.
Why can a fluent answer be wrong?
As our guide to how LLMs work explains, language models generate sequences of tokens using learned patterns. NIST notes that this can produce both accurate statements and inconsistent or inaccurate ones. Fluency is therefore not a built-in certificate of truth.
Training and evaluation also matter. Research published in Nature in April 2026 examines how accuracy-focused scoring can reward guessing over admitting uncertainty. If both a wrong answer and an honest abstention earn zero, guessing offers a chance of receiving credit. The authors also explain why sparsely represented facts pose problems even under idealized assumptions about training data.
That does not mean models are never trained to express uncertainty. Later training can explicitly encourage more reliable behaviour. The point is that different objectives and evaluations create different incentives; “the AI is never rewarded for saying it does not know” is too sweeping.
A chatbot application may also search the web, consult documents or use a calculator. It would be misleading to say none of these systems ever looks anything up. Access to tools changes the workflow, though the generated interpretation still needs checking.
The errors you are likely to meet
Imagine asking for a short history of a fictional town called Whiskerton. If you request a story, an invented mayor is part of the entertainment. If you request municipal records and receive Mayor Felix Purrington, elected in 1847, the same creativity has wandered into the wrong department.
More ordinary failures can be subtler:
- A reference-shaped invention: an impressive title, plausible authors and a journal volume that do not correspond to a real paper.
- A source mismatch: a real link is attached to a statement that the linked page does not support.
- A summary that goes too far: “the proposal will be discussed” becomes “the proposal was approved.”
- Mixed-up details: facts about two people or product versions are combined into one description.
Arithmetic and reasoning mistakes can accompany these failures, although not everyone classifies every wrong calculation as a hallucination. For the person using the answer, the useful question is usually how to verify it, rather than which filing cabinet the error belongs in.
A real case of imaginary cases
In Mata v. Avianca’s June 2023 sanctions order, the court described lawyers submitting nonexistent opinions generated by ChatGPT and continuing to defend them after their authenticity was challenged. A $5,000 penalty was imposed jointly on the respondents, alongside other requirements.
The lesson is broader than “a chatbot made a mistake.” The court identified failures of human verification and subsequent conduct. A formatted citation had been treated as evidence that a decision existed. Asking the same generator to reassure you is not a substitute for locating the decision itself.
What actually helps?
Providing relevant documents can improve factual work. The original retrieval-augmented generation research combined generation with retrieved material and found more factual outputs than its comparison model on the evaluated generation tasks. Our RAG explainer walks through the idea.
That is not a promise about every system carrying the label. A retrieval process can select the wrong document or an outdated passage. A model can then misunderstand what it received. Evidence needs to be relevant and the answer faithful to it.
For everyday use, make the checking process concrete:
- Identify the claims that matter. A playful headline and a quoted return deadline deserve different levels of scrutiny.
- Open the actual source. Check the author, date, relevant passage and whether it supports the specific claim.
- Separate extraction from interpretation. Ask what the document explicitly says and which conclusions are inferences.
- Verify numbers independently. Recalculate a total rather than accepting the paragraph explaining it.
- Leave gaps visible. “Not established by this source” is more useful than an invented answer that fills the space neatly.
Our prompt engineering guide offers examples of clearer briefs. Such instructions are aids to a workflow, not commands that make mistakes impossible.
Can another AI catch the problem?
Sometimes. A 2024 Nature study on semantic entropy investigated detecting a particular subset of hallucinations by examining variation in the meanings of generated answers. It is a research approach to uncertainty, not a universal truth detector.
For a reader, conflicting answers are a reason to investigate. Matching answers are not proof: systems can repeat the same error. Rephrasing a question can expose instability, but it cannot tell you which version is correct without evidence.
There is also no sound basis for promising that every newer model hallucinates less on every task, or for confidently predicting that the problem can never be solved. Measure the application you are using and the work you need it to do.
Enjoy the brainstorming, the drafts and the unusually patient explanations. When the answer depends on a real fact, follow it back to something checkable. The cat may remain certain about the corner; you are still allowed to turn on the light.
