An AI hallucination is an output containing unsupported or incorrect information that appears to be established. It can include invented facts, references, quotations or details about real subjects.
The problem is particularly difficult when the wording is fluent and plausible. Presentation quality is not evidence that a claim has been verified.
What can hallucinations look like?
A draft may introduce a product feature that was never supplied, invent a source or attach an incorrect figure to a real organisation.
Not every error has the same cause. Missing context, ambiguous instructions and source problems can interact with model behaviour. The practical task is to detect and address the unsupported result.
Why does confidence mislead?
Language models can produce confident wording without a corresponding verification process. A polished explanation may conceal uncertainty.
Do not use tone as a substitute for evidence. Ask what source establishes the important claims and whether that source actually supports the wording.
Does retrieval solve the problem?
Retrieved information can help ground a response, but the system can retrieve the wrong material or misrepresent a correct passage.
Evaluate both the retrieval and final answer. A citation is useful only when it supports the claim beside it.
How can workflows reduce risk?
Provide suitable reference information, define boundaries and make missing evidence an acceptable result. Use validation appropriate to the task.
For consequential content, review the claims before use. A general instruction to “be accurate” does not replace a checking process.
How should teams test for it?
Include incomplete, conflicting and unfamiliar cases in evaluations. Check whether the system recognises uncertainty rather than filling every gap.
Track the types of unsupported output and the effort needed to correct them. That evidence helps identify whether a change improves the actual workflow.
What should happen when an error is found?
Correct the result and investigate where it arose. Review the sources, instructions and any validation that should have caught it.
Avoid assuming that one prompt edit permanently solves the whole category. Re-test relevant cases and keep the operating limits visible.
What should businesses expect?
Expect a need for evaluation proportionate to the task. A private concept exercise and a technical customer instruction require different levels of confidence.
The useful goal is a process that recognises and manages uncertainty. AI output should not gain authority simply because it is easy to produce.
Further reading
For reference, see the supporting documentation. Explore related explanations of Generative AI, Retrieval-augmented generation, Prompt engineering.
A practical example
Illustrative example: a generated product description adds an unsupported runtime claim. The editor checks the supplied specification, removes the claim and adds that case to the evaluation set. The team then tests whether the workflow identifies missing specifications correctly.
Why it matters
Understanding hallucination helps a business separate plausible output from supported information. Appropriate evidence and review make the process more reliable.
