An LLM hallucination is a plausible-sounding answer that includes information that is false or unsupported. It can be a made-up source, an incorrect detail, or a confident conclusion that does not follow from the evidence. The examples below are invented teaching examples, not quotations from a real model. They show what to check when reviewing any generated answer.
OpenAI describes hallucinations as plausible but false statements generated by language models. A smooth tone does not verify a claim. Check the underlying source, the exact detail, and whether the answer distinguishes what is known from what is uncertain.
Example 1: A citation that looks real but does not exist
Invented answer: “The 2021 paper ‘Urban Canopy Effects on Sleep Quality’ by M. Rivera and L. Chen found that adding street trees increases average sleep by 42 minutes.”
What to check: Search the exact title and authors in the journal or a scholarly index. Open the paper itself. Confirm that it exists, that it studied the stated population, and that it reports the number in the answer. A citation format, DOI-shaped string, or confident author list is not evidence by itself.
Why this can mislead: Bibliographic details often follow predictable patterns, so an invented reference can look convincing. Search for the title and DOI independently instead of relying on a model-provided link.
Example 2: A real source with a claim it does not support
Invented answer: “A city report says transit ridership rose 18%, proving that the new fare policy caused the increase.”
The report may really say ridership rose 18%, but that fact alone does not establish why. Check the original report’s date, comparison period, measurement, and discussion of other changes. The model may have turned a correlation into a causal explanation.
Example 3: A fabricated detail about an obscure fact
Invented answer: “The 1912 edition of a small regional handbook lists the bridge as 640 feet long.”
Specific numbers and dates are easy to repeat and hard to notice when they are wrong. Look for the original handbook, scan, or a reliable archival record. If the source is unavailable, describe the figure as unverified rather than repeating it as fact.
Example 4: Contradicting the material in the prompt
Imagine that a passage says a fictional city has 80,000 residents in 2020. A response then states that the passage reports 120,000 residents. This is a contradiction that can be caught by comparing each claim with the supplied material. Ask the model to quote the supporting sentence, then verify that the quotation matches the original text.
Example 5: A correct answer attached to the wrong explanation
A response may reach the right result using a faulty reason. For example, it may identify the correct month in a schedule but claim that the event happens every three weeks when the listed dates show a monthly pattern. Check both the conclusion and the steps or evidence used to reach it.
A quick checklist for checking LLM answers
- Separate claims: Break a long answer into individual facts, dates, figures, and causal statements.
- Open the cited sources: Confirm the source exists and says what the answer claims it says.
- Prefer primary evidence: Check original documents, official data, research papers, or first-party records when available.
- Verify numbers and dates independently: Recalculate or compare them against the underlying table or record.
- Look for missing uncertainty: If evidence is incomplete, ask for caveats and inspect the source rather than treating confidence as proof.
- Compare the answer with the input: This helps catch omissions and contradictions when the model is summarizing supplied material.
Retrieval can make source checking easier, but it does not guarantee a correct answer. A model may miss a relevant passage, misread it, or add information that the documents do not support. Our guide to retrieval-augmented generation explains how retrieval supplies context and where it can fall short. The available context also matters; see what an LLM context window is.
Frequently asked questions
What is a simple example of an LLM hallucination?
A fabricated citation is a clear example: the answer names a plausible paper, author, or DOI that cannot be found, or cites a real paper that does not support the claim.
Can an AI hallucination sound confident and still be wrong?
Yes. Fluent wording and confidence do not prove a statement is true. OpenAI’s explanation of hallucinations notes that models can produce plausible but false statements.
Does retrieval-augmented generation prevent hallucinations?
No. Retrieval can provide useful evidence, but a model can miss or misread relevant text, or make claims that the retrieved sources do not support.
How can I tell whether an AI citation is real?
Search the title, author, and identifier independently, then open the source. Verify the exact passage, result, date, or figure before relying on it.
Source
- OpenAI: Why language models hallucinate — definition and explanation of plausible but false model outputs.