Why models make things up

A precise meaning for an everyday problem

In AI discussions, a hallucination usually means generated information that is unsupported or incorrect in the relevant context. It might be an invented quotation, a false citation, or a summary that adds a claim absent from its source.

Not every error has exactly the same cause. A model may lack relevant information, misinterpret the request, continue a false premise, or use retrieved material incorrectly. The important learner skill is to detect the unsupported step and repair the process.

Watch a claim appear from nowhere

Here is our entire invented source:

The library's Saturday workshop begins at 10:00. Registration is required. The notice does not name an instructor.

Request: “Summarise this workshop and tell me who teaches it.”

Bad answer: “The workshop starts at 10:00 on Saturday and is taught by Dr Lee.” The first part follows the notice. The named instructor does not. A sentence can mix supported and unsupported claims, so checking only the opening is not enough.

A better answer is: “It starts at 10:00 on Saturday, registration is required, and the notice does not identify the instructor.” Missing information stays missing.

Improve the request and the checking

Ask: “Using only the notice, answer each question and say ‘not stated’ when the answer is absent.” This gives a clear evidence boundary. It can improve behaviour, but the system may still make a mistake. Compare the answer with the notice yourself.

If the instructor matters, find an appropriate additional source, such as the library's current event listing. Do not ask the same model to invent a reference for its first answer. If it has browsing tools, verify that the source it returns actually names the instructor for this event.

Why fluent text can outrun evidence

The model produces text using learned patterns and the context supplied. A familiar question invites a familiar answer shape: a date, a place, a named expert. Producing that shape is not the same as possessing evidence for every detail inside it.

Modern systems may be trained to abstain or use tools. Those protections help but are not perfect. Avoid the equally misleading claim that every model is incapable of expressing uncertainty. What matters is whether this answer is supported.

Requests that deserve extra attention

Be especially careful with exact quotations, precise numbers, local events, recent changes, obscure people, and citations. A citation can look convincing because it has an author, year, and title while failing to identify a real source or a relevant passage.

Another trap is a question with a false premise: “Why did the library cancel its Saturday workshop?” If no cancellation occurred, an answer that explains the cancellation is already on the wrong path. First establish whether the premise is true.

Follow a missing-information failure from start to finish

Suppose you provide three meeting notes: “Garden club meets Thursday. Bring gloves. The notice does not name a room.” You ask for a complete invitation, and the answer says, “Join us Thursday in Room 12 at 4 p.m.”

The model has filled two gaps with plausible details. The output sounds like a normal invitation, but the room and time were not supplied. Neither normal wording nor a realistic room number makes them established facts. A system can produce the familiar shape of an invitation while failing the factual constraints of this invitation.

Change the request to: “Write an invitation using only these notes. Use [time needed] and [room needed] where information is missing.” A suitable result preserves Thursday and gloves and leaves the two gaps visible. Check that it actually did so; an instruction reduces ambiguity but does not guarantee obedience.

The next human action is to obtain the missing information from the organiser or an authoritative notice. Asking the same model for a more specific guess does not create that evidence. This is how you keep a useful drafting task from becoming an accidental source of invented facts.

Repair the result

Mark unsupported claims, remove or verify them, and distinguish facts from guesses in the final wording. Preserve uncertainty rather than filling every gap. This is useful even without AI: responsible writing does not pretend that every question has been answered by the available evidence.

Diagnose a made-up citation without relying on its appearance

An illustrative answer cites “Journal of Everyday Robotics, volume 18, pages 24 to 31” for a claim about classroom robots. A journal name, volume, page range, and confident sentence form a plausible pattern. Those parts do not prove that the article exists or that it supports the claim.

Break the problem into two checks. First, does a source matching the title, authors, date, and publication exist? Second, does its actual content support this precise claim? A real article about university robotics does not automatically support a statistic about primary-school classrooms.

If this is only a planning exercise, you can remove the unsupported statistic and write the argument without it. If the statistic matters, obtain an appropriate source before using it. Do not ask the model to repeat the invented citation more confidently and treat that repetition as new evidence.