An AI hallucination is an incorrect or misleading output presented by an AI system. In a chatbot, it might be a plausible explanation containing a made-up reference, an invented detail in a document summary or a link to a page that does not exist. A fluent answer can still contain an error, so the wording alone cannot establish whether the claim is supported.
Google Cloud's official explainer describes hallucinations as incorrect or misleading results and identifies problems with training data, mistaken assumptions and lack of grounding among possible causes. This guide uses that definition, with original hypothetical examples for students and everyday users. It does not assign an accuracy score to any assistant or claim that one product is immune.
| Output to check | What could be wrong | Useful next step |
|---|---|---|
| A named report | The report or its title may be invented | Find it on the issuing body's website |
| A document summary | It may include a detail absent from the source | Compare each key claim with the passage |
| A citation link | It may be broken or irrelevant | Open it and read the supporting text |
| A current deadline | It may repeat an older notice | Check the current official notice and date |
What does an AI hallucination look like?
Imagine asking an assistant to summarise a fictional college notice. The notice says applications open on Monday but gives no closing date. If the summary adds “applications close on Friday”, the new detail is unsupported. This is a hypothetical illustration, not a report about any college or an observed test of a particular model.
A second hypothetical case is a bibliography containing a convincing paper title, author and publication year. Those details can look complete while the cited work does not exist. Google Cloud specifically notes that AI systems can fabricate links to nonexistent web pages. Formatting a reference correctly is therefore different from establishing that the reference is real.
Hallucinations can also be quieter than a completely invented source. A summary may attach the right number to the wrong group, omit a condition or combine separate passages into an unsupported conclusion. The checking task is to ask whether the source supports the particular sentence, rather than merely whether the source mentions the general subject.
Why do AI hallucinations happen?
Google Cloud explains that AI models learn patterns from training data and use those patterns to make predictions. If the data is incomplete, biased or otherwise flawed, the resulting patterns can lead to incorrect outputs. An answer can sound coherent while reflecting an unreliable pattern rather than a verified fact.
The explainer also identifies a lack of grounding. Here, grounding means connecting a response to relevant factual material. Without a reliable connection to the document, database or public source needed for a task, a model can produce something plausible but incorrect, irrelevant or nonsensical. These are possible causes, not a diagnosis of why a particular response failed.
For readers, that distinction matters. You usually cannot inspect all the training data or determine the internal cause of an individual error. You can inspect the material used to support the answer. A practical review therefore focuses on claims, sources and missing information that are visible to you.
Does grounding or web search stop hallucinations?
Google's Gemini API documentation on grounding with Google Search says the feature connects the model to real-time web content and enables verifiable source citations. It describes grounding as a way to reduce hallucinations and improve factual accuracy. That is a stated purpose of the feature, not proof that every grounded answer is correct.
The difference is useful when choosing a workflow. A source-connected answer gives you material to inspect. You still need to open the citation, confirm that it is the right issuer and check whether the cited passage supports the claim. Our RAG explainer describes another common way of supplying retrieved material to an AI system.
Google Cloud's official guide to generative AI security terms recommends fact-checking and human oversight, especially for high-risk outputs. A search result or citation should therefore start the verification process. It should not be treated as a substitute for reading the relevant source.
How can I check an AI answer?
First, separate the answer into checkable claims: dates, names, amounts, eligibility conditions and quotations. Open the primary source for each important item. For a document task, compare the summary with the actual paragraph or table. If a detail is missing, leave it missing until an appropriate source confirms it.
An original prompt for source-bound summaries is: “Use only the document below. List each stated deadline with the exact supporting sentence. If a deadline is absent, write 'not stated'.” Its purpose is to make unsupported additions easier to spot. It is an instruction, not a guarantee against errors. Check the quoted passages yourself.
For reviewing a draft answer, try: “Identify every factual claim in this draft that needs an external source. Do not invent sources. Separate confirmed information from questions I still need to check.” This original prompt creates a review checklist. For more on structuring requests, see our prompt-writing framework and guide to checking cited web answers.
