AI fact checker: how to check every claim in a draft before it's published

This is our pillar guide to fact-checking with AI. For the step-by-step check of chatbot output, see how to fact-check AI-generated content. For the full editorial process, see how to fact-check an article.

How AI fact checking works: claim extraction, retrieval, matching, confidence

Fact-checking is the process of verifying every checkable statement in a piece, before or after it is published, against the best available source. AI tools split that process into four steps:

1. Claim extraction. The model reads your draft and pulls out sentences that make a checkable assertion: "Unemployment fell to [X]% last month." It skips opinions ("This is a bad policy") and questions. Good tools split compound sentences into separate claims.
2. Retrieval. For each claim, the tool searches a source collection: the web, a news archive, an index of scholarly papers, or documents you uploaded. The collection decides the ceiling. A tool that searches only the web will miss a paywalled study; a tool that searches only papers will miss a city council agenda.
3. Matching. The model compares the claim with what the retrieved sources say. Is the source about the same thing? Does it support, contradict, or say something close but different ("fell to [X]%" vs. "fell to [X]% among adults over 25")?
4. Verdict and confidence. The tool labels each claim, usually something like supported, contradicted, partly supported or no source found, and shows the source.

The weak point is step 4's last label. "No source found" is not "false." It may mean the claim is invented. It may also mean the source is a phone call, a local record, a paywalled database or an event from this morning. Only a human can tell which.

Which kinds of claims AI fact checkers handle well

The pattern: AI fact checkers are strongest where the truth is already written down somewhere public and specific. They are weakest where journalism adds the most value.

Step by step: run a claim check on a 1,200-word draft

Here's the process on a typical 1,200-word piece, using any claim-checking tool. Expect 20 to 40 checkable claims at that length.

Step 1: freeze the draft

Check the version you'll publish. If the text changes after checking, the changed sentences need checking again. Save a version (or note the version number) before you start.

Step 2: run the claim check

Paste or open the draft in your tool and run the check. You should get a list of extracted claims, each with a label and a source.

Check the extraction first. Skim the original and make sure every number, name, date, quote and finding is on the list. Tools miss claims buried in long sentences, captions and pull quotes. Add any that are missing.

Step 3: sort the results into three piles

Don't skip opening the "supported" sources. A tool can match a claim to a source on a similar topic that doesn't really support it.

Step 4: check the sources, not just the claims

A supported claim is only as good as its source. For every study, check that it hasn't been retracted, whether it's peer-reviewed or a preprint, and that the DOI points to the paper you meant. More in the next section.

Step 5: record what you checked

Keep a fact sheet: claim, source, status, who checked it. Here is a fictional example of three rows:

| # | Claim (para) | Source | Status | Checked by |
|---|--------------|--------|--------|------------|
| 1 | "55% of GPT-3.5 citations were fabricated" (p3) | Walters & Wilder 2023, Sci Rep, doi:10.1038/s41598-023-41032-5, abstract | Supported | JD |
| 2 | "The council voted 5-2" (p6) | Council minutes, 14 Oct, p. 3 | Supported | JD |
| 3 | "Ridership doubled since 2020" (p8) | No source found by tool; transit agency data shows +62% | Corrected to "rose 62%" | JD |

Our editor's fact-checking checklist has a full fact-sheet template.

Source quality matters as much as the claim: retracted papers, preprints, wrong DOIs

A claim can be "supported" by a source you shouldn't use. Three problems come up most often with research:

• Retracted papers. Retracted studies keep getting cited for years. The Retraction Watch Database, now owned by Crossref and free to use, lists tens of thousands of retractions.[1][2] Search the paper's title or DOI there before you rely on it.
• Preprints. A preprint hasn't been peer-reviewed. That doesn't make it wrong, but your sentence should say so ("a study that has not yet been peer-reviewed"). Check whether a later peer-reviewed version changed the findings.
• Wrong DOIs. AI-generated citations often attach a real-looking DOI to the wrong paper. Paste the DOI after https://doi.org/ and confirm the title, authors and year match. In one study of ChatGPT references, many fabricated citations came with plausible author names and journal formats, which is why they are easy to miss.[3]

AI fact checkers compared

Prices as of November 2026, from each vendor's site. Check before you buy.

Notes on each:

• Originality.ai is best known for AI-text and plagiarism detection, and the fact checker sits in the same scan. Its help center is candid that highlights are prompts for review, not verdicts.[4]
• Facticity.AI is made by AI Seer, a Singapore company, and stands out for checking claims in video. It publishes a 98.33% accuracy figure from its own benchmark; there is no independent test of that number.[6]
• Perplexity isn't a draft checker, but it's a fast way to research one claim, because every answer cites sources you can click. Check that each cited page says what the answer says.
• Google Fact Check Explorer only finds claims someone has already checked, using the ClaimReview markup that fact-checkers add to their articles. For viral claims, it's the right first stop. For your own original reporting, it will usually return nothing.[7]
• Grammarly Citation Finder is a good fit if your team already writes in Grammarly docs and wants sources suggested as you go.

Run a claim check in WriteOrbit

WriteOrbit's AI review checks the draft where you write it. Run a claim check and each claim gets a comment in the margin with what the tool found and the source it matched. Run a source-quality check and it flags citations to retracted papers, preprints and DOIs that don't match the paper. Because results are ordinary comments, your editor sees them, replies to them and resolves them, so the comment thread becomes your fact-check log.

It works best for drafts that cite research: science and health stories, reports, white papers. Citations in WriteOrbit come from an index of 250M+ scholarly works, so each one points to a real paper. If your sources are documents rather than papers, upload them to the Library and use Ask on each PDF; answers carry page numbers you can check.

What it doesn't do: verify quotes from your interviews, check images, or search published fact checks. Use the recording, a reverse image search and Google Fact Check Explorer for those. The free plan includes 3 review checks a day; see pricing and how it works.

What still needs a human

No AI fact checker replaces these checks. Budget your human time here:

• Quotes. Check every quote against the recording or your notes. Call back if you're unsure of wording.
• Anonymous and confidential sources. Only the reporter and editor can judge whether a source is in a position to know.
• Local facts. Council minutes, court dockets and local records often aren't indexed. Pull the document.
• Fast-moving news. For anything from the last 48 hours, rely on primary sources and direct contact.
• Context. A number can be true and misleading. Check the base year, the comparison and the definition. "Crime doubled" from 2 incidents to 4 is true and meaningless.
• Images and video. Use reverse image search and metadata tools; a caption claim is a claim too.
• What's missing. No tool tells you which important fact you left out.

The CNET case is the reminder: an audit found errors in 41 of 77 AI-assisted explainers, after editors had reviewed them.[9] Fluent text makes errors easy to read past. Check the claim, not the sentence.

FAQ

Is there a free AI fact checker?

Yes. Google Fact Check Explorer is free and searches fact checks already published by fact-checking organizations. Originality.ai's free plan includes its fact checker with limited credits, Facticity.AI lets you start for free, and WriteOrbit's free plan includes three AI review checks a day, including claim checks. Free tiers are enough for occasional drafts, not a daily newsroom load.

Can AI fact check itself?

Not reliably. Asking the same chatbot that wrote a claim whether it is true often gets the same error repeated with more confidence. A useful check needs an outside source: retrieval from a real index of papers, web pages or your own documents, and then a human who opens that source. Treat a model's self-check as a hint about what to verify.

How accurate are AI fact checkers?

There is no independent, standard benchmark for AI fact checkers yet, and vendors' accuracy figures come from their own tests. Accuracy depends on the claim type: they do well on published numbers, dates and scientific findings with good sources, and poorly on local facts, recent events, anonymous sourcing and context. Use them to find what to check, not to decide.

What is the best AI fact checker?

It depends on what you check. For a viral claim, start with Google Fact Check Explorer. For a whole draft, use a tool that extracts each claim and shows the source it matched, such as WriteOrbit's claim check or Originality.ai's Fact and Hallucination Checker. For research-heavy writing, pick one that also checks source quality, such as retracted papers and preprints.

Can an AI fact checker detect AI hallucinations?

It can catch many of them, especially invented citations and numbers that don't match any source. It catches them by failing to find support, so it can't distinguish an invented fact from a real one it simply couldn't find. A claim marked "no source found" needs a human to search further before you cut it or keep it.

Where to go next

If your draft started as chatbot output, follow the seven steps in how to fact-check AI-generated content. For the full editorial process with a printable checklist and free tools, read how to fact-check an article. Newsroom readers can see where checking fits in the journalist's AI workflow, and if you're choosing a writing tool, start with the best AI writing assistant for real sources.

Sources

1. Retraction Watch (2023). The Retraction Watch Database becomes completely open, and RW becomes far more sustainable. https://retractionwatch.com/2023/09/12/the-retraction-watch-database-becomes-completely-open-and-rw-becomes-far-more-sustainable
2. Crossref. Retraction Watch data documentation. https://www.crossref.org/documentation/retrieve-metadata/retraction-watch/
3. Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, 14045. https://doi.org/10.1038/s41598-023-41032-5
4. Originality.ai Help Center. Fact Checker. https://help.originality.ai/en/article/fact-checker-dqwo21/
5. Originality.ai. Pricing. https://originality.ai/pricing
6. Facticity.AI. https://www.facticity.ai/
7. Google News Initiative. Google Fact Check Tools. https://newsinitiative.withgoogle.com/resources/lessons/google-fact-check-tools/ ; Fact Check Explorer: https://toolbox.google.com/factcheck/explorer
8. Grammarly. Citation Finder. https://www.grammarly.com/ai-agents/citation-finder
9. Engadget (2023). CNET had to correct most of its AI-written articles. https://www.engadget.com/cnet-corrected-41-of-its-77-ai-written-articles-201519489.html