The short answer
To check if AI is telling the truth, split the answer into separate claims, ask for a source or exact quote behind each one, open those sources yourself and confirm the claims that matter in a primary source or a second tool. Anthropic's documentation says even its most advanced Claude models can state false information, and that grounding answers in quotes reduces but does not remove this. Perplexity, which cites every answer, and Gemini Notebook, which answers only from your files, make checking faster.
Why AI answers can be wrong even when they sound sure
AI assistants such as ChatGPT, Claude and Gemini write by predicting likely text, not by looking facts up in a database. Most of the time that produces correct answers. Sometimes it produces a hallucination: a fluent, confident statement that is false, such as a wrong date, a misquoted figure or a source that does not exist. Anthropic's own documentation says plainly that even its most advanced models can produce text that is factually incorrect or inconsistent with the material they were given, and that its techniques reduce this without eliminating it.
The tone of an answer tells you nothing about its accuracy. The good news is that checking is a skill with a short, repeatable routine, and the tools now give you more ways to do it than a year ago.
What you need first
- The answer and the chat it came from, so you can ask follow-up questions in context.
- A sense of the stakes. A brainstorm needs no checking. A number going into a board pack, a legal clause or a health question needs every claim checked.
- Access to a tool that shows sources, such as Perplexity, which lists numbered sources with every answer, or ChatGPT, Claude and Gemini with web search or research turned on.
- The original document, if the answer was about a file you uploaded.
How to check an AI answer in 7 steps
- Split the answer into claims. Ask: "List every factual claim in your answer as a numbered list: names, numbers, dates, quotes and causes." You cannot check a paragraph, but you can check a list.
- Mark the claims that matter. Usually only three or four claims carry the decision. Check those first and properly.
- Ask for a source or a quote for each one. For web facts, ask for the link. For your own files, ask for the exact sentence. Anthropic recommends making answers auditable by asking for a quote or source behind each claim, and retracting any claim that has none.
- Open the source yourself. A link proves nothing until you read it. Check that the page exists, that it says what the AI claims, that it is the original source rather than a copy, and that it is recent enough.
- Look for the primary source. For prices, features and policies, that is the company's own pricing or help page. For statistics, the study or official dataset, not an article about it.
- Ask a second, independent tool. Paste the claim, not the whole answer, into a different assistant and ask whether it is accurate, with sources. Disagreement is a red flag worth investigating. Our guide to comparing AI answers shows how.
- Remove or rewrite what you cannot confirm. If a claim has no source you trust, cut it or say plainly that it is unconfirmed. This is the rule we follow on every Yuzoki page.
Red flags that an AI answer needs checking
- Precise numbers with no source, such as "73% of managers" with nothing behind it.
- Quotes attributed to real people. Invented or paraphrased quotes are a known failure.
- Citations you cannot open, or titles that look plausible but do not exist when you search for them.
- Anything recent. A model's training data stops at a cutoff date, so new prices, product names and laws are risky unless the answer used live search.
- Answers about your document that you cannot find in your document.
- Different answers when you ask again. Anthropic lists running the same prompt several times and comparing results as a way to spot hallucinations.
A worked example: checking a statistic before it goes in a deck
You ask an assistant for a statistic on how many employees use AI at work, and it replies with a percentage, a year and the name of a well-known consultancy. It looks perfect for slide three.
- You ask: "Give me the exact title and link of the report this figure comes from, and quote the sentence that contains it."
- The assistant gives a link. You open it. The report exists, but the figure on the page is for a single country and a different year.
- You paste the corrected claim into a second tool with web search and ask whether a more recent figure exists from the same publisher.
- You use the figure from the original report, with its real scope ("in the UK, in [year]") and a footnote link.
Five minutes of checking turned a confident error into a claim nobody in the room can knock down. That is the habit worth building.
Common mistakes when checking AI and how to fix them
- Asking the same chat "are you sure?" The model may simply agree or apologise without checking anything. Fix: ask for a source or quote, then open it yourself.
- Treating a link as proof. Fix: read the cited page and find the sentence.
- Checking the easy claims and skipping the important one. Fix: rank claims by what breaks if they are wrong.
- Using a model without search for current facts. Fix: turn on web search or use a tool built around sources, then check the date of each source.
- Relying on secondary write-ups. Fix: go to the maker's page, the original study or the official register.
Which AI tool to use for fact-checking
- Perplexity is the most convenient checker for web facts, because every answer lists numbered sources. On the Max plan, Model Council runs one question across three models and shows where they agree and disagree. Note that Free, Pro and Max accounts can be used for training unless you switch off AI data retention in Preferences. See how to use Perplexity.
- Gemini Notebook from Google (called NotebookLM until 16 July 2026) is best for checking claims about your own documents, because it answers only from the sources you add, with citations. See our Gemini Notebook guide.
- Claude is strong at quote-checking long files: ask it to quote the passage behind each point. Research, which reads many sources, starts on the Pro plan.
- ChatGPT includes a limited allowance of deep research even on the free plan, which writes cited reports. See how to use ChatGPT deep research.
For research-heavy work, our best AI for research page compares these in depth, and ChatGPT vs Perplexity covers the two most common choices. To make fewer errors in the first place, read how to reduce AI hallucinations.
4 prompts to copy for checking AI answers
Use these in a new chat or a second tool, so the checker is not defending its own earlier answer.
Prompts to copy
Break an answer into checkable claims
Here is an AI-written answer: [paste]. List every factual claim in it as a numbered list. For each claim, say whether it is a number, date, name, quote or cause, and rate how much it would matter if it were wrong (high, medium, low). Do not judge whether the claims are true yet.
You get: A ranked checklist of what to verify
Source every claim or drop it
For each numbered claim below, find a primary source (the company, study or official body that published it), give the link and quote the exact sentence that supports it. If you cannot find a supporting source, write NOT FOUND rather than guessing. [paste claims]
You get: Sources and quotes you can open, with gaps clearly marked
Check a summary against the document
I have attached [the document] and pasted a summary of it below. For each sentence of the summary, quote the passage in the document that supports it. Mark any sentence that the document does not support, or supports only partly, and explain the difference. Summary: [paste]
You get: A line-by-line audit of a summary
Second opinion on one claim
Is this claim accurate as of [month year]? "[claim]". Search for the most recent primary source, give the link and its date, and say whether the claim is correct, outdated, partly correct or wrong. If sources disagree, show both.
You get: An independent check with dated sources