What Is AI Hallucination, and How Do You Reduce It?

Why does AI hallucinate?#

AI tools hallucinate because a language model generates text by predicting what is likely to come next, based on patterns in its training data. It does not look facts up unless a tool supplies them. The same prediction that often produces accurate text can also produce text that is fluent, confident and wrong.

NIST's risk profile for generative AI uses the term confabulation for this, and notes that it is colloquially called hallucination or fabrication. NIST describes it as a natural result of the way generative models are designed: they produce outputs that approximate the statistical patterns of their training data, for example by predicting the next token. A 2023 research survey on arXiv puts it simply: large language models (LLMs) are prone to generating plausible yet nonfactual content.

NIST adds that the problem is particularly relevant for open-ended prompts asking for long answers, and in areas that need deep context or specialist expertise. That matches everyday experience: a short summary of a page you supplied is less risky than a long answer about a specialist topic from memory.

If you are new to how these models work, start with our plain-English guide to LLMs.

What does hallucination look like at work?#

Hallucination at work usually looks like a wrong fact stated with confidence, an answer that drifts from the document you supplied, an invented reference or explanation, or a statement that contradicts something said earlier. It rarely announces itself. The wording is usually as polished as a correct answer, which is what makes it risky.

IMDA's Starter Kit for Testing LLM-Based Applications describes the risk as a tendency to produce incorrect or fabricated output with respect to universal facts or specific sources or documents, such as company policy. NIST also includes outputs that diverge from the prompt or other input, and outputs that contradict earlier statements in the same context.

NIST warns that confabulated logic or citations can make people trust an answer more than they should, because the output appears to justify itself. A list of neat references is not evidence until you have opened them.

How do you reduce AI hallucination?#

You reduce hallucination by giving the tool reliable source material, allowing it to say it does not know, asking for supporting quotes, and checking important claims yourself. Anthropic's guidance on reducing hallucinations sets out these basic techniques, plus a few advanced ones, and is clear that they reduce the problem without eliminating it.

The techniques below come from that guidance, written for everyday use.

Anthropic recommends the quotes-first technique especially for long documents, and describes the citation step as having the tool find a supporting quote for each claim after it answers, and retract the claim if it cannot.

Grounding answers in your own documents#

Supplying source material is called grounding. Some tools do this automatically using retrieval-augmented generation (RAG), which, in Anthropic's glossary, retrieves relevant information from a knowledge base and passes it to the model with your question. The glossary notes that RAG's effectiveness depends on the quality and relevance of what is retrieved, and the arXiv survey discusses the limits retrieval-augmented models still face. Grounding helps; it does not replace checking.

Matching checks to the stakes#

IMDA's starter kit notes that each app needs its own threshold. Hallucination in a poetry app would likely have low impact, but in a medical chatbot or immigration screening tool it could cause real harm. Apply the same thinking to your own tasks: the more a wrong answer could hurt someone, the more human in the loop checking it needs.

What might this look like in a Singapore workplace?#

Picture a twelve-person events company in Bugis that wants a short FAQ for new hires, drawn from its staff handbook. Asked from memory, an AI tool might describe a leave or claims rule that sounds standard but is not in the handbook. Grounding the request in the handbook text, and checking each answer, addresses this.

This is an illustrative scenario. A careful approach could look like this:

1. Paste the relevant handbook sections. Use the tool your company has approved, and leave out anything containing staff names or personal details.
2. Restrict and allow uncertainty. "Answer only from the handbook text below. If the handbook does not cover a question, write: not covered, ask HR."
3. Ask for quotes. "Under each answer, include the handbook sentence it is based on."
4. Check every answer. Someone who knows the handbook compares each answer with its quote, and removes or fixes any that do not match.
5. Record what changed. Note any corrections, so the same prompt can be improved next time.

The "not covered" answers are useful too. They show where the handbook itself may need updating.

What are the common mistakes?#

The most common mistakes are asking for facts from memory, trusting confident wording, not opening the references the tool provides, and treating one technique as a complete fix. Most come from forgetting that a language model produces likely text, not checked facts, so every important claim needs a source you trust.

• Asking for facts from memory. Supply the source instead, or check every fact against one.
• Treating confidence as accuracy. NIST notes that the confident nature of a response can lead people to believe and act on false content.
• Not opening the references. An invented reference can look exactly like a real one.
• Asking open-ended questions on specialist topics. Narrow the question and supply the material.
• Overloading one conversation. Very long chats can fill the context window and bury your instructions. Our guide to tokens and context windows explains why.
• Assuming the problem is solved. Anthropic's guidance says its techniques significantly reduce hallucinations but do not eliminate them, and that critical information should always be validated.

For a structured way to test AI output before others rely on it, see how to test and evaluate AI-generated outputs.

What to learn next#

The next step after understanding hallucination is learning how the model's working memory, measured in tokens, affects what it can keep track of. From there, better prompting habits and a simple testing routine will help you get more reliable results from AI tools in everyday work.

• Find more practical guides for everyday users in AI for users.
• Follow the AI learning pathway to see how this topic links to prompting, building and governance.
• To practise grounding, testing and checking on a real project, the GenAI Builder Sprint is a 2-day classroom course where adults plan, build, test and launch a simple AI app. No previous AI or programming experience is needed, although formal entry requirements apply.

Sources#

1. Reduce hallucinations, Anthropic (Claude Platform Docs), accessed 8 November 2026
2. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, National Institute of Standards and Technology (NIST), accessed 8 November 2026
3. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions, arXiv, accessed 8 November 2026
4. Starter Kit for Testing LLM-Based Applications for Safety and Reliability, Infocomm Media Development Authority (IMDA), accessed 8 November 2026
5. Glossary, Anthropic (Claude Platform Docs), accessed 8 November 2026

Frequently asked questions#

What is AI hallucination?+

It is when an AI tool produces content that sounds confident and plausible but is false, invented or not supported by the material it was given. NIST calls this confabulation and notes that it is colloquially known as hallucination or fabrication.

Why do AI tools hallucinate?+

Language models generate text by predicting the next token based on patterns in their training data. NIST describes confabulation as a natural result of that design: the same statistical prediction that often produces accurate text can also produce false or inconsistent text.

What forms does hallucination take?+

Common forms are wrong facts stated as true, answers that do not match the document you supplied, invented references or reasoning, and statements that contradict something the tool said earlier in the same conversation.

How can I reduce hallucination in my prompts?+

Give the tool the source material and tell it to use only that. Allow it to say it does not know. For long documents, ask for word-for-word quotes first, then base the answer on them. Ask it to support each claim with a quote and remove claims it cannot support.

Can hallucination be removed completely?+

No. Anthropic's guidance says these techniques significantly reduce hallucinations but do not eliminate them, and that critical information should always be validated, especially for high-stakes decisions.

How much checking does an AI answer need?+

It depends on the harm a wrong answer could cause. IMDA's starter kit notes that hallucination in a poetry app would likely have low impact, while hallucination in a medical chatbot or immigration screening tool could cause real harm.