Why AI Hallucinates And Gets Facts Wrong

A museum assistant says that the archive reading room will be open until 16:00 on 12 June 2026. It cites a real supplied notice. The notice gives those hours for the temporary gallery and says the archive reading room is closed that day. What must we check beyond

the existence of a citation? Follow the failure, the supporting passage, the response it permits and the evidence needed to evaluate uncertainty. The example is fictional, but the distinction can be inspected directly. Name the specific failure. Here the system retrieved a relevant

notice but transferred the gallery's hours to the archive and missed its closure. That is a source-interpretation error. An invented notice, a wrong calculation, invalid output formatting and an unauthorized booking would be different failures. Calling every problem hallucination would hide which component needs repair.

Hallucination commonly describes unsupported or incorrect generated content under a task. A model can fill an absent detail with a plausible continuation, repeat a stale fact or misinterpret supplied evidence. In our case the required closure statement is already in the notice, so adding another citation is

not the first missing operation. The claim must be matched to the actual passage's facility, date and exception. Check grounding at the claim level. The citation points to notice C2, but the sentence about the reading room needs support from the part about that room. Read the scope and conditions,

not just the document title. The supported response to this supplied case is that the reading room is closed on 12 June. The gallery's separate opening interval does not override that statement. Change the task slightly: a visitor asks about another date

that the notice does not cover. The workflow now lacks applicable evidence. It can ask for the missing date, retrieve an authoritative current notice, state the available limit, abstain or route the case to staff, depending on what is missing. The response should help resolve the actual gap rather than merely append a

generic warning to an unsupported opening time. If the application can reserve a room or book a visit, uncertainty must affect that action path. Hold the action while its required evidence or permission is missing. A footer saying the answer may be wrong does not repair an

automatic booking based on unverified hours. The application needs defined conditions for proceeding, clarifying or escalating, with a record of what evidence justified the decision. Model knowledge can be incomplete, stale or poorly elicited. A stated model knowledge cutoff concerns the model's reported training boundary; a product may

also use current retrieved information or tools. Inspect what the actual request received and which source applies. Neither a cutoff date nor the presence of a search tool establishes that the current answer used correct, authorized evidence. TruthfulQA supplies a dated evaluation case. Its 2021 preprint and May 2022 ACL

paper used questions designed around human misconceptions to test whether models repeated false answers. The study separated truthfulness from informativeness. This illustrates a useful evaluation distinction: a detailed-looking answer and a truthful useful answer need separate checks. Its historical models and deliberately constructed questions do not provide a present-day

error rate for the museum system. Confidence also needs a defined meaning. Suppose a hypothetical evaluation groups 100 answers assigned an 80% correctness score, but only 60 are correct under the stated rubric. Their observed correctness is 60%, below the assigned score. That bin reveals overstatement in this

supplied example; a complete calibration assessment needs adequate representative data and its protocol. The word certainly is not by itself an evaluated probability, and a next-token probability does not automatically measure whole-answer truth. Keep retrieval success and evidence-supported correctness as separate outcomes. Notice C2 can be retrieved successfully

while the answer still misreads it. Also inspect whether a correct answer was formatted properly and whether any action respected permission. These distinctions let the team repair retrieval, interpretation, validation or action controls instead of assuming that changing model size addresses every failure.

Pause with the supplied notice: the temporary gallery is open from 10:00 to 16:00 on 12 June 2026, and the archive reading room is closed that day. The assistant cites it while claiming the reading room is open until 16:00. Did retrieval succeed? Does the passage support that claim?

State the response justified by these givens. The notice was retrieved, but it does not support the reading-room opening claim. The hours belong to the temporary gallery, while the reading room is explicitly closed on the specified date. State the reading-room closure and preserve the facility and date distinction.

This is a source-interpretation repair, not evidence that any cited sentence is automatically grounded. Try another domain. A supplied bus timetable says route 7 runs from 07:00 to 19:00 on weekdays, from 10:00 to 14:00 on Saturday, and has no Sunday service. A generated answer cites the timetable

and offers to book a Sunday trip at 11:00 using the weekday hours. Identify the failure and the supported response. What should happen to the booking action? The answer has applied the wrong day condition. The timetable says there is no Sunday service, so those weekday hours

do not support the proposed trip. Explain the supplied Sunday limitation and hold that booking route. An alternative trip would require applicable evidence and permission. A real citation and confident wording cannot authorize a trip that contradicts the cited conditions.

You can now name a failure, inspect whether a passage supports a particular claim, and choose a response that changes the workflow when evidence is missing or contradictory. Useful uncertainty is specific and consequential. The final lesson will turn these distinctions into a controlled comparison, so that model capability and changes in system

input are evaluated without being mixed together.