How to Fact-Check With AI: Verify Claims, Sources and Websites

When we ran a real, still-circulating claim through this process, "Scotland generates more than 100% of its electricity from renewables", every source an AI assistant offered checked out: an official government release, a trade body, Wikipedia. The claim was still misleading. That gap is exactly why AI fact checking works best when the model helps investigate evidence instead of standing in for it, and it cuts in both directions: whether AI is helping you check a claim or produced the claim in the first place, the evidence has to come from sources outside the model.

Below is a repeatable workflow for checking claims, sources, websites and AI-generated answers, worked through step by step on that exact claim, plus the prompts that go with each stage.

What Is AI Fact Checking?

AI fact checking is the use of artificial intelligence to help identify factual claims, find relevant evidence, compare sources and flag inconsistencies during the verification process. AI can accelerate the research, but the final judgment should rest on the underlying sources rather than the model’s answer alone.

People searching for this usually mean one of two things. Fact-checking with AI is using a model to check a statistic, investigate a website or gather evidence. Fact-checking AI is verifying something a chatbot produced: a citation, a study reference, a generated number.

Both come down to the same task of reaching the actual source and reading what it says.

The Limits of AI as a Fact-Checker

Yes, it can help, though not as an autonomous truth detector.

Language models can generate fluent answers that are not reliably factual, especially when information is missing, recent or highly specific. The failure modes are catalogued further down; the short version is that a model can produce a confident answer, a real citation and a wrong conclusion in the same paragraph.

Professional practice reflects this. Full Fact, which has been building machine learning into fact-checking since 2016, designs its tools to find claims worth checking rather than to decide whether a claim is true or false, and says plainly that it doesn’t treat AI as a cure-all or expect human experts to become unnecessary. Research on automated verification agrees, since many claims are partially correct, or correct but misleading without context, and resist binary labels.

One line to carry through the rest of this guide: a sourced answer is not automatically a verified answer.

How to Fact-Check With AI: A 7-Step Workflow

Copy-ready prompts for these steps are collected further down.

1. Turn the Statement Into a Checkable Claim

Verification starts with knowing exactly what’s being tested. "This article is misleading" can’t be checked. "The article says US remote workers are 35% more productive than office workers" can. If a paragraph contains several claims, split them.

Whatever prompt you use, tell the model not to verify anything yet. Asking for a verdict at this stage invites one before any evidence exists.

2. Ask What Evidence Would Verify the Claim

This step separates a real workflow from typing "is this true?" Before searching, work out what a good answer would look like. Which original source should exist? What data would settle it? Which organizations hold authoritative figures, and which dates, populations or geographies change the meaning?

For an employment statistic, that points you at the Bureau of Labor Statistics and the underlying survey instead of ten blogs quoting the number.

3. Find the Original Source

If the claim rests on a study, statistic, survey, report, quote, law, filing or government dataset, go to the original. AI can help locate it; you have to open it.

Check the title, author, date, methodology, sample and the wording of the relevant passage or table. An AI summary describes the evidence. The document is the evidence.

4. Check Whether the Source Actually Supports the Claim

A citation can exist, be genuinely reputable, and still fail to support the specific claim attached to it. A paper reporting a 40% improvement might be reporting it for one subgroup, as a correlation, on self-reported data, or on a different metric entirely.

Ask the model to point at the exact evidence and name what the claim overstates. Then open that passage and read it yourself.

5. Find Independent Evidence

One source is a starting point. Look for another study, government data, a different credible organization, independent reporting or a subject-specific database.

Two links aren’t necessarily two independent sources. Article A cites Article B, which cites a press release, which describes a company’s own study — one evidence chain wearing four hats. When you ask a model for corroboration, tell it to exclude pages that trace back to the same origin, because it won’t work that out on its own.

6. Check Who Is Behind the Website

For an unfamiliar site, look into the publisher, the organization, the author’s expertise, ownership, funding where relevant, and any correction history. A useful technique here is lateral reading: leaving the original website and checking what independent sources say about it.

7. Write a Verdict That Matches the Evidence

Most real claims don’t resolve to TRUE or FALSE. A scale that fits actual evidence runs from supported through mostly supported, partly supported, misleading and unsupported to contradicted or unclear.

Record the claim, the verdict, the primary evidence, the independent evidence, the context that matters, and what remains uncertain. Restrict the model to the sources you actually checked, and tell it not to fill gaps with assumptions.

A Real AI Fact-Checking Example

Here’s the workflow run end to end on a claim that circulates widely in energy coverage.

Claim: Scotland generates more than 100% of its electricity from renewable sources.

Step 1: What the AI returned first

Asked with web search enabled, an AI assistant confirmed the claim and cited three sources: a Scottish Government press release, the trade body Scottish Renewables, and Wikipedia. Nothing hallucinated. The lead citation was an official government page.

Step 2: What the sources actually said

Opening them changed the sentence. The Scottish Government’s January 2024 release says renewable technologies generated the equivalent of 113% of Scotland’s overall electricity consumption in 2022, the highest figure recorded. Scottish Renewables reports the same number.

The load-bearing word is equivalent. The government’s own methodology page defines the measure as all renewable electricity generated within Scotland divided by gross electricity consumption, which is total generation minus net exports. Scotland exports a large share of what it generates, which shrinks the denominator. The figure is a ratio between two national totals, not the renewable share of the electricity Scots actually use.

Step 3: Independent evidence

Two sources with no stake in the government’s framing addressed this directly. The UK Statistics Authority, asked to review the claim, explained that as a gross measure it does not mean all electricity used in Scotland came from renewables, and that in the year to September 2022 the renewable share of electricity actually consumed in Scotland was 63.1%. Full Fact reached the same conclusion, noting Scotland produces more electricity than it uses, including a substantial amount from fossil fuels and nuclear.

Step 4: Check the date

The 113% figure is from 2022. More recent government statistics show that of the 51.8 TWh generated in Scotland in 2024, 73.1% came from renewable sources and 91.5% from low carbon sources including nuclear and pumped storage.

Verdict: partly supported, misleading as usually phrased

The number is real, official and correctly quoted. The wording wrapped around it is where it breaks. "The equivalent of 113% of gross electricity consumption" is accurate; "more than 100% of its electricity" implies something the statistic doesn’t measure, and the renewable share of electricity actually consumed in Scotland was around 63% for the period the statistics authority examined.

The AI found the right sources in seconds and took a reader straight past the distinction. Reading the government’s own definition of the metric is what resolved the claim.

How to Verify AI Sources and Citations

When an assistant gives you a citation, run four checks in order.

1. Does the source exist? Confirm the URL resolves and that the title, authors, publication and DOI correspond to something real. Crossref, Google Scholar, Semantic Scholar and the publisher’s own page are the fastest routes, plus PubMed for biomedical work.

2. Does the metadata match? Models frequently attach a real paper to the wrong year, journal or author list, or blend two papers into one plausible reference.

3. Does the source support this specific claim? The most important of the four and the one most often skipped. An existing citation and a supporting citation are different things.

4. Is it the original source? If you’ve landed on something that describes research rather than reporting it, keep going until you reach the research.

Being indexed in a database confirms a paper exists. It says nothing about the quality of the research or whether its findings apply to the claim in front of you.

How to Fact-Check a Website With AI

Start by getting the page to declare itself.

List the main factual claims on this page and the evidence the page provides for each one. Separate evidence from opinion or marketing language.

Then run steps 3 to 6 on whatever it cites, and check the publisher against sources other than its own About page.

The part worth adding here is how you frame the conclusion. Skip the verdict on the whole domain, because "is this site trustworthy?" is close to unanswerable and rarely what you need. Ask instead whether this specific page is reliable enough for the specific claim you want to use. A solid publication can carry a weak article, and a promotional site can accurately quote a real dataset.

How to Fact-Check ChatGPT or Other AI Answers

The same workflow applies to chatbot output, with two habits that are specific to it.

Don’t verify a model by asking it "are you sure?" It will often restate the same error more confidently, and self-correction isn’t independent evidence. And check anything time-sensitive separately, because prices, laws, software features, company leadership and policies all drift — a correct 2023 fact can be a wrong 2026 answer.

Prioritise the claims that carry weight: numbers, dates, quotes, studies, legal and technical specifics, and anything that changes month to month.

Prompts for AI Fact Checking

Five prompts covering the workflow above. They work as a sequence, and each one assumes you’ll open what it returns.

Extract claims

Extract every independently checkable factual claim from this text. Ignore opinions and predictions. Put the claims in a table and explain what evidence would be needed to verify each one.

Find primary evidence

Find the original or primary source for this claim. Prefer official data, the underlying research paper, original report or direct record over articles that cite it.

Check source support

Does this source actually support the claim below? Identify the exact evidence, explain any important limitations, and flag anything the claim exaggerates or leaves out.

Find independent evidence

Find independent sources that confirm, contradict or add context to this claim. Exclude pages that all trace back to the same original source.

Build the evidence table

Create a fact-check table with Claim, Evidence For, Evidence Against, Primary Source, Important Context, Confidence and What Still Needs Verification. Do not make a binary judgment where the evidence is uncertain.

Common AI Fact-Checking Mistakes

1. Asking "is this true?" Too vague to produce evidence, and it pushes the model toward a verdict it has no basis for.

2. Trusting citations that look legitimate. Formatting, DOIs and journal names are easy to generate.

3. Asking the same model to verify itself. Useful for surfacing internal inconsistencies, worthless as independent confirmation.

4. Treating agreement between models as proof. If ChatGPT, Claude and Gemini say the same thing, that’s not three sources. They may draw on similar web material or reproduce the same widespread misconception. Model consensus is a signal, and evidence is what settles it.

5. Counting repeated reporting as independent evidence. Check where each page got the claim before counting it.

6. Ignoring dates. The Scotland example turns partly on which year’s statistics you’re reading.

7. Flattening nuanced evidence into True/False. Some claims are accurate in one framing and misleading in another, which a binary label erases.

What AI Fact Checkers Can and Cannot Do

AI Fact Checker Tools vs. AI-Assisted Fact Checking

There are two ways to bring AI into verification.

A dedicated AI fact checker takes a claim, URL or document and returns an automated analysis, often with a confidence score. Fast, simple, a reasonable first pass. Its methodology is usually opaque, a second model can reproduce the same mistake as the first, and a confidence score tells you how confident a system is without telling you whether the evidence behind it is any good.

An AI-assisted manual workflow uses AI to find, organize and compare while you open the evidence yourself. Slower, and transparent in a way the first approach isn’t.

For low-stakes checks, an AI fact checker is a useful filter. For anything you plan to publish, cite or act on, use the second approach.

Useful Tools for Fact-Checking With AI

Google Fact Check Explorer surfaces fact checks published by other organizations. Google doesn’t produce or endorse the checks themselves.

How Sigma Browser Fits an AI Fact-Checking Workflow

Sigma isn’t a fact checker, and no browser can tell you whether a website is true. What it does is keep the page, the AI assistance and the wider research in one place.

Start on the page itself. Chat with Page will pull out what needs checking if you ask it directly:

List the factual claims on this page that would need independent verification. For each claim, show what source the page provides, if any.

From there the work moves outward. Open the cited documents rather than reading the summary of them, and use Deep Research when a claim has been covered across many sources instead of one. Comparing what comes back — claim against evidence, methodology, date and contradictions — is where most of the time goes, and there’s a walkthrough of that in this guide to summarizing and comparing multiple web pages.

Then come back to the original page and decide: supported, overstated, misleading, unsupported, or still unclear. Sigma cuts the copy-pasting and tab-switching involved in getting there. The judgment at the end is still yours.

When AI Is Not Enough

Some checks shouldn’t end with a general-purpose model. Medical decisions, legal interpretation, significant financial choices and specialist technical questions need authoritative current sources and, where it matters, qualified professional input.

Images and video are their own problem. An AI detector reporting "92% likely synthetic" is a probability estimate from a system with its own error rate, and it isn’t proof. Media verification usually depends on reverse image search, metadata, the earliest available upload, geolocation, timestamps and independent footage. Detecting whether media was AI-generated and checking whether a claim is factually accurate are also separate questions with separate methods.

AI Fact-Checking Checklist

Before you trust an AI-assisted fact check:

1. Isolate the exact claim.
2. Open every important citation.
3. Check that the source supports that specific claim.
4. Trace statistics, studies and quotes to the original.
5. Find genuinely independent evidence.
6. Check dates and context.
7. Separate uncertainty from contradiction.
8. Base the conclusion on evidence, not on the model’s confidence.

Should You Use AI for Fact Checking?

Yes, as a research assistant.

AI meaningfully speeds up claim extraction, source discovery, reading long documents and comparing what several sources say. In the Scotland example it found the right government page in seconds. What it missed was that the word "equivalent" carried the whole claim, and that’s what decided the verdict.

If your checks usually involve jumping between an article, its sources, a few PDFs and a dozen tabs, Sigma lets you run more of that inside the browser you’re already reading in.