What you’ll be able to do

By the end of this, you’ll be able to take a claim, a post, a headline, whatever’s bugging you, and get an AI tool to hand you back actual sources you can click. Not a vibe. Not “trust me”. Real links, real dates, stuff you can check yourself. That’s the whole point, right, because the AI’s answer is only ever the starting point, never the final word.

Before you start

Fact-checking matters more than ever because so much of what lands in front of us has already been stripped of its context. A number gets pulled out of a report, a quote gets clipped from a longer speech, a graph gets cropped, and suddenly it’s everywhere. You’ve probably shared something like this yourself without checking. Most of us have.

Here’s the bit people get wrong about AI though: it doesn’t “know” if something’s true. It’s not connected to some master database of facts sitting there waiting to be consulted. What it can do, if you give it web search access, is go and look things up in real time, pull in sources, and summarise what it finds. That’s genuinely useful. But it’s also where things go wrong, because these tools can and do make up sources that look completely real. Fake links, fake studies, fake quotes attributed to real people. Not because the AI is being sneaky, it just predicts what a source-shaped answer would look like, and sometimes that prediction is wrong.

So what you’re about to learn isn’t “ask AI, get truth”. It’s a method. Find the original source. Check the date. Check the number actually matches. Ask the assistant to argue against itself. That’s the skill, not the tool.

Get set up

For this job you want an AI assistant with web search switched on, or a tool built specifically for search with sources attached.

Perplexity is built for exactly this. Every answer comes with clickable citations sitting right next to the claims, so you’re not hunting for where something came from. ChatGPT, Claude and Gemini can all search the web too, when that feature’s turned on, and they’ll usually show links alongside what they tell you. Copilot does this as well through its search integration. Worth checking that web search is actually enabled in whichever one you’re using, because some will answer confidently from memory alone if you don’t nudge them.

Alongside the AI, there are tools built purely for fact-checking that don’t use AI at all in the way people expect, and they’re worth having bookmarked. Google Fact Check Explorer is free and searches published fact-checks from outlets around the world. Full Fact, based in the UK, and Snopes both do dedicated fact-checking work and are worth searching directly for anything that’s been circulating for a while. If the “fact” in question is a photo, reverse image search with Google Lens or TinEye will often show you where an image actually came from and when, which kills a lot of misleading claims stone dead. And if you’re on X, Community Notes sometimes already has context attached to a post, so check there first before you do anything else.

Before you start typing, have the actual claim in front of you, word for word if you can. Not your memory of it. The exact wording, the source it came from if you know it, and roughly when you saw it. That’s your starting kit.

Try it yourself

Here are three prompts, each doing a different job.

For checking a specific factual claim, statistic or quote:

Fact-check this claim: "[paste the exact claim here]"

I saw this in [where you saw it, e.g. a social media post, a news article, a speech].

Please:
1. Tell me whether this claim is accurate, misleading, or false, and why
2. Give me the original source of the underlying data or statement, not just articles about it
3. List the specific sources you used, with links, so I can check them myself
4. Tell me if the number/date/context has changed since the original source was published

For checking a whole article or post that feels off:

I want you to help me fact-check this article, not summarise it.

Here's the text: [paste the article or a link if you have web search on]

Please identify:
1. Any specific claims, statistics or quotes that could be checked against a primary source
2. For each one, whether it's stated accurately compared to what you can find
3. Any claims that are unverifiable or that no source supports
4. What would need to be true for the article's main argument to hold up

For stress-testing an AI’s own answer, which is the step almost nobody does:

You just told me [paste the AI's previous answer or claim here].

Now argue the opposite case as convincingly as you can. What evidence, if it existed, would prove this claim false? Is there any of that evidence available right now?

That last one sounds odd the first time you use it, but it’s genuinely the most useful of the three. It forces the AI to stop being agreeable and actually stress-test its own output.

Check the result

Right, so you’ve got an answer back. Now the actual work starts, because the AI’s answer is not the fact-check. It’s a lead.

Click every single link it gives you. Every one. Does the page exist? Does it say what the AI claimed it says? Assistants will summarise a source in a way that’s technically defensible but leaves out the one detail that changes everything, and they will also cite pages that don’t exist at all. You won’t know which one you’re dealing with until you click.

Check the date on the original source against the date of the claim you’re checking. A statistic from 2019 doing the rounds as though it’s current is one of the most common ways misinformation spreads. Numbers age. Context ages faster.

Check that it’s actually the same number. This sounds obvious but it’s where a lot of misleading claims live. “Crime is up 40%” might be technically sourced from a real report, but if that report measured a different crime category, a different area, or a different time period than the one being claimed, it’s not really the same fact anymore.

Try searching the claim independently through Google Fact Check Explorer or by searching Full Fact and Snopes directly. If something’s been widely shared and it’s false, there’s a decent chance someone’s already written it up properly, and that write-up will usually be more thorough than anything an AI generates on the spot.

And if there’s an image involved, run it through Google Lens or TinEye separately. A photo can be entirely real and still be lying to you about when or where it was taken.

If it doesn’t work

Sometimes the AI just gets it wrong, and it’s worth knowing what that looks like so you’re not caught out.

The biggest one: fabricated sources. The assistant gives you a link, a study name, even a page number, and none of it exists, or it exists but says something different. This isn’t rare. Treat every citation as unverified until you’ve clicked it yourself.

Watch out for confident vagueness too, answers that sound thorough but don’t actually point anywhere specific. “Multiple studies show” with no studies named is a red flag, not a fact.

A screenshot of a headline is not a source. Neither is a quote tweet, a forwarded WhatsApp message, or “someone I trust said”. If the AI’s answer relies on you having sent it one of these as proof, push back and ask for the primary document, the actual paper, the filing, the transcript.

On privacy: don’t paste anything into an AI chat that you wouldn’t be comfortable with existing somewhere outside your control. That includes screenshots with personal details in them, anything involving someone’s medical, legal or financial situation, internal work documents, or messages from private conversations. Fact-checking a public claim is fine. Fact-checking someone’s private messages, or feeding in identifying details about a real person to “check” a rumour about them, isn’t something an AI tool is the right place for, and in some cases it isn’t something you should be doing at all.

Keep exploring

If this has been useful, a few other guides on this site build on the same ground. How to get AI to be honest with me goes into why these tools hedge, flatter, or agree with you when they shouldn’t. How to get AI to help me research a topic widens this out from single claims to proper research sessions. And How to get AI to find news on companies applies the same checking habits to business claims and press coverage specifically.

Sources and review notes

Review date: 15 September 2026 — every source above was opened and checked against the text on that date.