Hallucination

Definition: A hallucination is an AI assistant's output that presents false information as fact, contradicts supplied material, or conflicts with its earlier statements. The US National Institute of Standards and Technology (NIST) describes this phenomenon as confabulation.

Key facts

• NIST's risk management profile, published in July 2024, includes confidently presented false content and contradictions with input in its definition of confabulation.
• NIST's discussion of confabulation identifies fabricated citations and faulty supporting logic as ways an answer can mislead readers.
• Grounding connects a model's output to verifiable information sources. Google Cloud's Grounding overview, updated 15 October 2026, describes this connection as reducing the chances of invented content.
• As of 17 October 2026, Anthropic's Reduce hallucinations guidance recommends allowing uncertainty and checking claims against supporting quotations.

Why do hallucinations occur?

A large language model (LLM) is a model that generates language by predicting the next word or other text unit in a sequence. NIST's explanation of confabulation connects the problem to this statistical prediction process: models generate outputs that approximate patterns in their training data. That process can produce both accurate statements and factual errors.

NIST highlights open-ended requests for long responses and tasks requiring contextual or specialist knowledge. It also explains why a fluent answer deserves scrutiny: confident presentation can lead readers to believe and repeat false information. An explanation of how an answer was reached can itself contain fabricated reasoning.

What does an unsupported claim look like?

Google Cloud's grounding check documentation, updated 13 October 2026, supplies this deliberately incorrect example:

Google was founded by Larry Page and Sergey Brin in 1975

The documentation identifies the founders' names as correct and the date as wrong. It therefore treats the whole claim as ungrounded. In that service, full grounding requires every claim to be supported by the supplied reference facts; partial support is insufficient.

This example illustrates a useful review principle for website owners: check each factual component of a statement about a business. A sentence can combine correct names with an incorrect date.

How can a website owner check an answer?

Anthropic's verification guidance recommends making each claim auditable through supporting quotations and sources. It also recommends permitting the assistant to acknowledge uncertainty.

A practical application is to open a cited page, locate the relevant passage, and compare its wording with the answer's claim. For an answer about a business, the review can record the statement beside its source URL and supporting text. Google Cloud's grounding example shows why that comparison needs to cover the complete statement. NIST's warning about fabricated citations makes the existence and content of the cited source part of the check.

Related terms

• Grounding
• Retrieval-augmented generation (RAG)
• AI citation
• Large language model (LLM)

FAQ

Can an answer with citations contain a hallucination?

Yes. NIST's confabulation discussion explicitly includes fabricated citations. Open the cited source and check whether its text supports the claim.

Does a confident explanation make an answer reliable?

NIST's risk profile describes both confidently presented errors and fabricated logic. Assess the factual support for the answer and its explanation.

What can I ask an assistant to do during verification?

Following Anthropic's guidance, ask it to supply a supporting quotation and source for each claim, and allow it to acknowledge uncertainty. Check those quotations against the original pages.

Sources

1. NIST, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile", July 2024
2. Google Cloud, "Grounding overview", updated 15 October 2026
3. Google Cloud, "Check grounding with RAG", updated 13 October 2026
4. Anthropic, "Reduce hallucinations", accessed 17 October 2026

---

Check how a page reads to answer engines. Open the GEO score analyzer