What are AI hallucinations, and how do you check an AI answer?
AI hallucinations sound plausible but can be false or inconsistent with supplied information. Learn to check claims, citations and document-based answers.
On this page
What is an AI hallucination?
An AI hallucination is an answer that sounds plausible but is false or inconsistent with the information supplied to the AI. A chatbot might invent a reference or add a detail that the document it is summarizing does not contain. Confident wording does not make the answer accurate. [1] [3] [2] [4]
To check an answer, examine its important claims individually. Open the cited source and find the passage supporting each claim. A real citation is only the beginning: the source must also support what the answer says. [3] [4]
What hallucinations can look like
The examples below are hypothetical illustrations, not observed chatbot responses or product tests. They show different ways an answer can invent information or go beyond its source, alongside a practical check adapted from vendor guidance. [1] [3] [2] [4]
| Hypothetical claim | Possible error | What to check |
|---|---|---|
| Your phone has a setting called Battery Repair. | Invented technical detail | Find the setting in official documentation for your exact phone model. [1] [3] [2] |
| The manual says, “Charging overnight doubles battery life.” | Fabricated quotation or reference | Confirm that the manual exists, then locate the exact quotation and its context. [2] [4] [3] |
| A summary says a device supports wireless charging, but the supplied document never mentions it. | Unsupported addition to a summary | Look for a passage supporting that feature; otherwise mark it unverified from this document. [1] [3] [4] |
Why AI produces plausible mistakes
A language model is the type of AI behind many chatbots. OpenAI explains that its initial training learns patterns by predicting text. That process does not attach a true-or-false label to every statement, so producing plausible wording is different from establishing that a statement is true. [1]
OpenAI’s research also argues that training and evaluation can reward guessing instead of acknowledging uncertainty. This helps explain why plausible false answers persist; it is one explanation, not a diagnosis of every incorrect response. [1]
Search tools can help a chatbot find more current and verifiable information. However, OpenAI still advises users to visit sources directly and check important claims. Having access to search does not replace checking the answer. [2]
A worked citation check: real page, unsupported claim
Imagine an AI answer says: “Anthropic guarantees that restricting answers to supplied documents eliminates hallucinations.” It cites Anthropic’s documentation page titled Reduce hallucinations. This is a hypothetical answer used to demonstrate a source check. [3] [4]
The page exists and recommends restricting answers to supplied documents. But it also warns that its techniques do not eliminate all hallucinations. Our interpretation is that the citation passes the existence check and fails the support check: the recommendation does not justify a guarantee. [3] [4]
A narrower statement is supported: “Anthropic recommends restricting answers to supplied documents as a way to reduce hallucinations, while warning that errors can remain.” The qualification changes what the reader can reasonably conclude. [3] [4]
How to check an AI answer
This procedure is an editorial synthesis of OpenAI and Anthropic guidance. It separates finding a source from checking its support, and preserves qualifications that an AI summary may leave out. [2] [3] [4]
- Isolate the claim. Identify the specific fact, quotation or technical instruction you need to verify. [2] [3]
- Open the cited source. Confirm its publisher and title rather than relying on the chatbot’s description of it. [2] [3] [4]
- Find the supporting passage. Check whether it establishes the exact claim, rather than merely discussing the same subject. [3] [4]
- Read the surrounding text. Look for exceptions, uncertainty and other qualifications omitted from the answer. [4]
- Check the scope. Where relevant, compare the source’s date, device model, software version and region with the claim you are checking. [2] [4] [3]
- Record what remains unresolved. If you cannot access the source or find support, mark the claim unverified. Missing support alone does not establish that the claim is false. [3] [4]
A prompt that keeps the source in view
When you have a document to work from, you can adapt this prompt: “Answer using only the document below. For each factual claim, identify the supporting section and a short passage. If the document does not provide the answer, say so. Separate documented facts from interpretation.” Then supply the document. [3] [4]
This prompt is an editorial adaptation of Anthropic’s guidance, not a tested accuracy guarantee. Check the passages yourself: asking for supporting text does not establish that the answer or quotation is correct. [2] [3]
Before relying on an important factual answer, open its sources and verify the claims that affect your decision. If a key claim remains unverified, keep that uncertainty visible rather than treating fluent wording as evidence. [2] [4] [3]
For background, read what tokens mean in AI and our guide to writing clearer AI prompts.
