Does AI Make Things Up? AI Hallucinations and 5 Ways to Fact-Check

AI sometimes describes books, laws and numbers that do not exist as if they were real. Here is why hallucinations happen and 5 checks anyone can do.
Key Takeaways
- NIST AI 600-1, published by the US National Institute of Standards and Technology in July 2024, calls AI's confident production of false content "confabulation," commonly known as hallucination.
- A September 2025 paper by OpenAI and Georgia Tech researchers argues that training and evaluation methods that reward guessing are one reason language models keep hallucinating.
- Anthropic's official documentation says techniques to reduce hallucinations cut errors significantly but cannot eliminate them, so important information still needs checking.
- Five ways to check AI answers: open the sources, go back to primary sources, check numbers and names, ask again differently, and consult professionals on important decisions.
Picture a hypothetical case: you ask an AI to recommend a few books, and it confidently lists titles with authors and publication years, only for you to discover that one of them does not exist. That is what people call an AI hallucination, and it happens even with the most advanced models.
Drawing on the US National Institute of Standards and Technology (NIST), an OpenAI research paper and official documentation from AI companies, this guide explains what AI hallucinations are, why they happen and which questions are most error-prone, then lists 5 fact-checking methods anyone can use.
What is an AI hallucination?
An AI hallucination is when an AI states something false or made up in a confident tone, such as inventing book titles, papers, legal provisions or figures that do not exist, while making them look entirely real.
NIST is the US government agency that develops technical standards. In NIST AI 600-1, its July 2024 risk management profile for generative AI, it calls this phenomenon "confabulation," defined as the production of confidently stated but erroneous or false content, colloquially known as "hallucinations" or "fabrications," by which users may be misled or deceived.
NIST also notes that confabulation goes beyond factual errors. It includes answers that drift from your question or contradict earlier statements, and even made-up logic or citations that seem to justify a wrong answer. The document points out that AI sometimes lays out plausible reasoning steps even when the answer itself is wrong, and because the tone is so confident, users are more likely to believe it and act on it.

Why does AI make things up?
Language models predict the next word from statistical patterns rather than looking things up in a database. Training and evaluation also tend to reward guessing, so when a model does not know an answer, it leans toward guessing instead of saying "I don't know."
NIST AI 600-1 explains that generative models produce outputs that approximate the statistical distribution of their training data; large language models, for example, predict the next word in a sentence. This kind of prediction often produces accurate content, but it can also produce content that is wrong or internally inconsistent.
The paper "Why Language Models Hallucinate," published on September 4, 2025 by OpenAI researchers and a Georgia Tech scholar, uses an exam analogy: if a correct multiple-choice answer earns 1 point and a blank or "I don't know" earns 0, guessing on questions you are unsure about gives a higher expected score. The paper argues that most current evaluations work like these exams, keeping models permanently in "test-taking mode," so they tend to give a specific answer even when uncertain.
The paper gives an example: researchers asked an advanced open-source model for the birthday of the paper's lead author, adding "if you know," and the model gave three different wrong dates in three attempts. The paper also explains that rarer facts are more error-prone: if 20% of birthday facts appear only once in the training data, a base model after pretraining would be expected to hallucinate on at least 20% of birthday questions.
Which questions are most likely to trigger AI hallucinations?
Obscure names and birthdays, book and paper citations, legal provision numbers, recent news, and arithmetic or counting are common trouble spots. Questions that need long answers or specialist knowledge also call for extra care.
| Error-prone type | Why to be careful |
|---|---|
| Obscure people, birthdays, niche events | The paper above notes that facts rarely seen in training data are more likely to be guessed wrong |
| Book, paper, court ruling and statute citations | NIST warns that AI may fabricate citations; a complete-looking reference does not mean it exists |
| Recent news, prices and rules | Google recommends enabling Grounding with Google Search when obscure or recent facts are needed |
| Arithmetic, counting, statistics | Google recommends letting Gemini run code whenever arithmetic or counting is involved |
| Long answers needing specialist knowledge | NIST notes that open-ended, long-form answers and highly specialized fields are especially prone to confabulation |
Even simple counting can go wrong. The same paper describes researchers asking how many Ds are in the word "DEEPSEEK": an advanced open-source model answered 2 or 3 in ten separate trials, and other models gave answers as high as 6 or 7, when the correct answer is 1.
5 Ways to Fact-Check AI Answers
1. Ask for sources, then open them yourself
Anthropic's official documentation suggests having the AI cite quotes and sources for each claim so its answer can be audited. But the links an AI provides can also be fabricated, so always open them: check that the page exists, that it really says what the AI claims and that the date is current.
2. Cross-check against primary sources
Compare what the AI says with the original source: government websites for policies, Taiwan's Laws & Regulations Database run by the Ministry of Justice for laws, the original text for research papers, and official company announcements for products and prices. For rumors about government policy, Taiwan's Executive Yuan website has a real-time news clarification section.

3. Check numbers, dates, names and quotes one by one
The OpenAI research paper notes that when models guess, they often give overconfident and highly specific answers, such as "September 30" rather than "sometime in autumn." The more specific a number or date, the more it needs checking, and you should redo any calculations yourself with a calculator or spreadsheet.
4. Rephrase the question or ask it several times
Anthropic's documentation suggests running the same prompt several times and comparing the outputs, since inconsistent answers can signal hallucinations. You can also follow up from a different angle, for example by asking the AI to explain its basis or list what it is unsure about.
5. Leave important decisions to professionals
NIST gives a healthcare example: a confabulated summary of patient information could lead doctors to an incorrect diagnosis. For health, legal, financial and contract matters, treat AI answers as reference only and confirm with a doctor, lawyer, accountant or the relevant authority.
How can you ask so that AI makes things up less often?
Allowing the AI to say "I don't know," supplying your own material and asking it to answer only from that material, and asking for direct quotes all reduce hallucinations, though none can prevent them completely.
- Let it say it does not know: Anthropic's documentation says explicitly permitting the AI to admit uncertainty can drastically reduce false information. Add a line such as "If you are unsure, just say you don't know" at the end of your prompt.
- Use only the material you provide: After pasting a document, tell the AI explicitly to use only that material and not its general knowledge.
- Extract quotes before analyzing: For long documents, first ask the AI to pull out the relevant passages word for word, then answer based on those extracts, which keeps it closer to the original text.
- Turn on search: For recent news, use a mode with web search, and remember to open the pages it cites.

For more prompting techniques, see these 7 prompt writing tips. Anthropic also cautions that these techniques greatly reduce hallucinations but do not eliminate them, so always confirm critical information. If you are choosing a tool, see our 2026 AI chatbot comparison; to spot fake videos and voices made with AI by scammers, read 6 telltale signs of AI scams and deepfakes.
Frequently Asked Questions
Do paid plans or web search stop AI from hallucinating?
No. The OpenAI research paper says hallucinations persist even in state-of-the-art models. Google recommends search for recent or obscure facts, but you should still open the web pages the AI cites and check them yourself.
Can I trust the reference links an AI gives me?
Not without checking. NIST warns that AI may fabricate citations, so a link may not exist or the page may say something different from the AI's claim. Check the address, title, publisher and date before relying on it.
Is an AI hallucination a deliberate lie?
It is more like guessing on an exam question you cannot answer. NIST describes confabulation as a natural result of generative models producing content by statistical prediction, and the OpenAI paper says scoring that rewards guessing pushes models to answer even when uncertain.
How does Taiwan's government view AI-generated content?
Generative AI guidelines issued by Taiwan's Executive Yuan in October 2023 say agencies must not fully trust AI-generated information, and must not use unconfirmed output as the sole basis for official decisions.
Sources: National Institute of Standards and Technology – NIST AI 600-1 Generative Artificial Intelligence Profile (2024), Kalai et al. – Why Language Models Hallucinate (2025), Anthropic – Reduce hallucinations, Google AI Studio – Prompt design strategies; also the draft guidelines attached to a press release from Taiwan's National Science and Technology Council dated August 31, 2023, and a notice on their formal issuance from the Ministry of Economic Affairs' Small and Medium Enterprise and Startup Administration.