Fact-checking with AI is starting to work
We’re releasing Big If True, a skill that turns Claude into a careful, claim-by-claim fact-checker. We tested it on articles with known errors, and it caught them. Grab it for free at verso.ink/big-if-true.
Be honest: who checks your work? A generation ago, at the big magazines, the answer was a person with your draft, a red pencil, a near fatal amount of coffee, and days to spend. Today, for most of us, the answer is you, at 11 p.m. In this world, mistakes happen.
For the past couple of years, asking an AI to catch them was a good way to add new ones. The models skimmed confidently, green-lit whatever garbage sounded plausible, and made up sources without blinking.
But with the latest frontier models, this is changing. All through 2026, we quietly built and tested a fact-checker. In the last few months, we watched it start to work in real time: with a strong model and a rigorous method, it began catching real errors in real published stories.
Today we’re releasing it. It’s called Big If True, a free skill for Claude. Give it an article, a draft, a newsletter, a press release, and it works through the text the way a careful fact-checker would: find every checkable claim, verify each one against sources, and deliver a verdict you can audit.
Every claim means every claim
A couple years ago at Stanford, we were told a fun story about how legendary fact-checking desks work: If a draft says “Jesus Christ, the son of God,” you underline it, grab a Bible, and do your best to figure it out!
As you can see, the bar for rigorous fact-checking is up in heaven. But of course, it’s a bar very few desks can afford to clear. Big If True is an earthly way of trying. It underlines every claim it can find, and each one gets one of three verdicts:
- Supported
- Two independent sources agree, or one definitive source settles it.
- Couldn’t verify
- Claude can’t prove it either way. Often a quote only you heard, or a document only you have.
- Needs review
- Something could be off here. Have a look.
Notice the wording: there is no “wrong” verdict. This is by design. The skill can tell you where to look, but the final call belongs to a human.
Where does the evidence come from?
For every claim, the skill looks for the best available record. Take finance reporting as an example: For a bond yield or an inflation figure, it uses code to look up the official data directly. A quote is checked against a transcript, a study against its journal record. Big If True works natively across languages, so a Spanish piece will be checked against Spanish sources first. And importantly, a news story simply repeating a claim doesn’t count: five outlets carrying the same wire copy are treated as one source.
Big If True is a safety net under your work, with receipts attached. The receipts live in a shareable report that carries the full article with every checked claim highlighted.
Digging into campaign filings
Sometimes the record you need isn’t on a page a search engine can hand you. Let’s say it sits in a PDF or even a hand-written form. That’s where Claude’s browser comes in: in Cowork, Claude can open a page, click, and type, the way you would.
Here’s an example from our tests. In one of the articles where we planted mistakes, we changed a donation by Sergey Brin to a California ballot committee from $9 million to $4 million. Every run without the skill had shrugged: couldn’t verify.
But once the skill was used, Claude opened the state’s campaign finance database in the browser, went to the search page, typed “Sergey Brin” into the contributor field, expanded the results to 100 rows, and opened the filing PDF. Verdict: needs review, with the filing as the source. This was, of course, the correct answer.
Does it work?
To find out, we built a gold-standard test: real published pieces with errors we had identified, and some edited pieces where we planted mistakes ourselves.
We then ran the whole range of frontier models against the evaluation dataset, improving the pipeline and prompts along the way. Step by step, the fact-checks got better.
Here’s one piece of data from our testing. The first step in fact-checking is extracting factual claims from a long article. In our evaluations, we checked how many claims Claude extracts with and without the skill using Opus 5 and Fable 5.1:
Claims per article (average)
Without the skill, Claude picks only the claims that it thinks are worth checking, averaging to about 29 in our tests. With the skill, it examines nearly a hundred! This extra coverage makes a difference: in every run without the skill, a significant number of claims went unexamined, and real errors in the article were never caught.
The skill is also more careful about what it flags. Across the same runs, we counted how often each version wrongly accused the text of a possible error:
False accusations per run (average)
Without the skill, Claude cried wolf more than once per run on average. With the skill, that happened about once in ten runs: not perfect, but a big improvement.
What we found
Beyond the test set, we ran the skill on plenty of real articles. We don’t want to single anyone out, but once the pipeline worked, we found mistakes in many of them, from highly renowned newsrooms to smaller outlets all around the world. There was a buffet of errors: revenue figures a year out of date, three numbers that didn’t sum to the total the sentence claimed, governors mistaken for senators, all the way to a forecast reported as a finding.
Of course, we pointed the skill at ourselves as well. We are both delighted and horrified to report that it saved us from embarrassment multiple times.
There are limits. When sources block AI access or aren’t online, Big If True has less evidence to work with. Claims based on your own interviews, notes, or private documents may come back as “couldn’t verify” unless you provide them. And it can still get things wrong, so give the results a look before making changes.
Try it on your own work
Big If True is free. Download it at verso.ink/big-if-true, add it to Claude, and run it on the next piece you’re about to publish. Everything you need to run it is explained on that page.
We’ll keep improving the skill. Subscribe to our newsletter and we’ll tell you when a new version ships.
Finally, our plug: If you want help bringing Claude and skills into your newsroom, hire us. We can teach you how to build this stuff yourself. We’re also interested in developing versions of the skill that connect directly to your CMS, so the checks run against your own archive.
And if Big If True catches something in your draft before your readers do, we’d love to hear about it: [email protected].