The short answer
To reduce AI hallucinations, give the model your source material, tell it to answer only from that material, allow it to say "I don't know", ask for exact quotes before the answer and have it check each claim against a source. These techniques come from Anthropic's guide to reducing hallucinations and Google's Gemini prompting guide. Turn on web search for recent facts and use code for calculations. Anthropic says they reduce errors but never eliminate them, so still check what matters.
What an AI hallucination is and why it happens
A hallucination is when an AI model states something false as if it were true: a wrong figure, an invented source, a clause your contract does not contain. It happens because language models generate likely text rather than retrieving verified facts, and because they are trained to be helpful, which can push them to fill a gap rather than admit it.
You cannot switch hallucinations off. Anthropic's guide to reducing them ends with a clear warning: the techniques significantly reduce hallucinations but do not eliminate them, so always validate critical information. What you can do is set up your prompts and tools so that errors are rarer and easier to catch. This page covers prevention. For checking an answer you already have, see how to check if AI is telling the truth.
What you need first
- The source material the answer should come from: the report, policy, data export or web pages.
- A tool that can read files or search the web, so the model works from material rather than memory.
- Five extra seconds per prompt for the grounding lines below.
How to reduce AI hallucinations in 8 steps
- Give it the material instead of asking it to remember. Upload the document or paste the data. OpenAI's prompt engineering guide recommends adding relevant context the model would not otherwise have, which developers call retrieval-augmented generation (RAG).
- Restrict it to that material. Say "Answer only from the attached documents, not from general knowledge." Anthropic calls this external knowledge restriction. Google's prompting guide suggests a strict grounding rule: if the answer is not explicitly in the context, say the information is not available.
- Give it permission to say "I don't know". Anthropic says explicitly allowing uncertainty can drastically reduce false information. Add: "If the material does not contain the answer, say so."
- Ask for quotes first, then the answer. For long documents, Anthropic recommends asking Claude to extract word-for-word quotes before doing the task, so the answer is built on the actual text.
- Ask it to verify its own claims against sources. After a draft, ask the model to find a supporting quote for each claim and remove any claim it cannot support. This is Anthropic's "verify with citations" technique.
- Turn on search for anything recent. Google's guide recommends Google Search grounding for recent or obscure facts. In the apps, that means web search, Deep Research or a source-first tool such as Perplexity.
- Let the tool calculate instead of guess. Google advises using code execution for arithmetic and counting. Ask: "Use code to calculate this and show the formula."
- Use a stronger model or more effort for hard questions. Bigger models and higher reasoning effort tend to make fewer errors on difficult prompts. OpenAI reported that GPT-6.1 Terra's error rate on its internal set of tricky prompts fell from 53.6% for GPT-5.6 Terra to 16.7% at low effort, and to 4.3% at maximum effort. Those are OpenAI's own numbers, but the direction is consistent: effort matters.
A worked example: summarising a policy without invented clauses
The risky prompt: "What does our travel policy say about expenses abroad?" with no file attached. The model will produce a plausible generic policy, and nothing in the answer tells you it is not yours.
The grounded prompt: attach the policy PDF and write: "Using only the attached travel policy, answer: what can employees claim for meals and hotels abroad? First, list the exact quotes that are relevant, numbered. Then answer using only those quotes, citing them by number. If the policy does not cover something, write 'Not covered in the policy' rather than guessing."
The answer now shows its evidence. If quote 3 does not say what the summary claims, you see it immediately. And if the policy is silent on, say, per diem rates, you find out before a colleague relies on a number that was never there.
Common mistakes that cause hallucinations and how to fix them
- Asking about recent events with search off. Fix: turn on search or use a research mode, and check source dates.
- Asking for "sources" without search. A model without web access may invent references. Fix: only trust citations from a tool that actually searched or from your own files.
- Uploading a file bigger than the plan can read. ChatGPT's free plan has a context window of about 27,000 tokens, roughly 12 pages of input, so it may only see part of a long PDF. Fix: split the file or use a tool with a larger context window, such as Google AI Pro's 1 million tokens.
- Leading questions. "Why did sales fall in March?" invites an explanation even if sales rose. Fix: ask "What happened to sales in March, according to the data?"
- Very long chats. Fix: start a fresh chat with the key material when a conversation has drifted.
- Using the fastest model for judgement work. Fix: see how to choose the right AI model.
Which AI tool hallucinates least for your task
The honest answer is that the setup matters more than the brand. Still, some tools are built to make grounding the default.
- For your own documents: Gemini Notebook (formerly NotebookLM). Answers are grounded only in your notebook sources, with citations, and the free Standard plan allows 100 notebooks with 50 sources each. Its catch: a notebook built from weak sources gives confident, well-cited answers that are still wrong. See our guide.
- For web facts: Perplexity, because every answer lists numbered sources. See how to use Perplexity.
- For long, careful reading: Claude, with the quote-first technique from Anthropic's own guide. See how to use Claude.
- For cited reports: deep research in ChatGPT or Gemini, both with a free allowance.
Comparing options? See best AI for research, best AI for summarising PDFs and Gemini Notebook vs Perplexity.
5 prompts that reduce hallucinations
Add these to any task where accuracy matters. They follow Anthropic's and Google's published techniques.
Prompts to copy
The grounding line
Answer only from the material I have provided below, not from general knowledge. If the material does not contain the answer, say "Not in the material" instead of guessing. <material> [paste or attach] </material> Question: [your question]
You get: Answers limited to your sources, with gaps admitted
Quotes first, then the answer
From the attached [document], first extract the exact quotes most relevant to [question] and number them. If there are none, write "No relevant quotes found." Then answer the question using only those quotes, citing them by number.
You get: An answer you can trace back to the text
Self-check a draft
Review the draft above. For each factual claim, find a direct quote from the attached sources that supports it. If you cannot find one, remove the claim and mark where it was with [ ]. Then list the removed claims.
You get: A cleaned draft and a list of unsupported claims
Calculate, don't estimate
Using the attached [spreadsheet], calculate [the metric] with code rather than estimating. Show the formula or code you used, the rows you included and the result. Flag any missing or odd values you found.
You get: Numbers with the working shown
Confidence labels
Answer [question]. After each claim, add a label: [from your sources] with the source, [general knowledge, unverified], or [uncertain]. Do not present uncertain claims as fact.
You get: An answer that shows which parts need checking