How we check things
A channel about wrong beliefs is worth nothing if it adds new ones. So here is exactly what happens to a claim before it reaches the board.
Every claim is checked before it is drawn
The order matters. Fact-checking happens after the script is written and before the voice-over is recorded. At that point a correction costs one edit. After recording it would cost a new voice-over, a new board and a new render. That pressure is exactly what makes people leave a shaky claim in.
Three verdicts, not two
- Verified: we found a primary source and it says what we say it says.
- Wrong: the claim does not survive. It goes out, or it becomes the myth.
- Unverified: we could not confirm it. This is a verdict of its own, and it is the default for anything we could not trace.
That third verdict is the one that matters. Plenty of repeated-everywhere facts have no primary source at all. Take the pleasing detail that a matador's cape is red to hide blood. We could dress a guess up as a fact and nobody would notice. Instead we say we could not verify it, out loud and in writing. A claim we cannot check is not a claim we present.
Unverified covers three different failures, so on the page we print which one. Not established is the plain case: we checked, and nobody could pin the claim down.No source means a checker did back the claim, but the paper it cited could not be found, and a citation nobody can open is not a citation. Second check disagreed means the first reading confirmed it and the second one would not. Both readings are described further down, and when they split we side with the doubt.
The checker is never the writer
A model asked to check its own text agrees with itself, so the check is done by a different one from the one that drafted the script, and disagreements go to a person rather than getting quietly rewritten. Silently "fixing" a fact means replacing the author's claim with a second guess, which is how errors get laundered into confidence.
Two checkers, and they do not get an equal vote
Every claim is read by two different models that never see each other's work. They fail in opposite directions, which is the entire reason there are two: one is quick to call something false when the honest answer is that nobody knows, and the other will not confirm even the obvious. Run alone they get fourteen and fifteen of our twenty test claims right. Run together they get eighteen.
So the votes are not weighed equally. Wrong needs both of them to say so, because that verdict prints a correction in the middle of an article, and an invented correction is worse than a missing one. Verified needs only one, because the claim still has to clear the source rules after that. When the two disagree, the claim stays unverified. We would rather publish a gap than a confident mistake.
The sources are published, not implied
"Studies show" is not a source. Every claim ships with the actual reference: author, venue, year, and a link where one exists. It goes on the video's page here, in the newsletter, and in the video description. If we cannot name it, we say so instead.
We will not invent a citation to look thorough. On a site about checking things, a fabricated source would be the worst thing we could publish.
Searched, or judged from knowledge
There is a difference between a claim someone went and looked up and a claim ruled on from what a reference work already says, and most fact-checking hides it. Ours prints it. Next to every verdict you will see searched ornot searched.
Both are legitimate, for different claims. That Jupiter is the largest planet does not need a search, and refusing to affirm it is not caution but a broken instrument. What the label protects you from is the third case: a confident verdict dressed up as research. If it says not searched, weigh it as what a well-read checker believes, and the source line tells you where that belief comes from.
We used to let the checkers search the web themselves and stopped, because we measured what came back. A checker with a search box returns whatever is at the top of the results: a pet shop, a listicle, a site that exists to rank. The same model working from memory named Kawchuk in PLOS ONE and Hyde in American Psychologist. Searching also made it worse at the simple things. It went to the web and came back unsure whether water boils at 100 °C. So the web now does one job it is actually good at: confirming that a paper the checker named exists. If we cannot find it, the source comes off and the claim drops to unverified, however plausible the citation looked.
The label is not the checker's word for it. Under the current setup searched means we took the source it named and went looking for that source on the web, so the label is a fact about a request we made, not a claim the model gets to volunteer. When a checker does have a search box, every query it runs is recorded on our side and the mark only survives if a real query matches the claim. Either way, a checker that claims credit for work it did not do gets corrected by the machinery rather than believed.
Older pages carry no label at all. Those checks happened before we started recording which line was looked up, and marking them either way would be a guess. They pick up labels when the page is next rebuilt.
The checker is measured, not trusted
A fact-checker is software, and software that nobody measures drifts. Ours runs against a fixed set of twenty claims with known answers, five of them corrupted on purpose, four of them familiar myths. The set is written outside the model that does the checking, because a test written by the thing it tests only proves they agree.
The number that matters is how many false claims come back "verified". It is zero, and every change to the checker has to keep it there. The current pair also catches all nine planted falsehoods outright, with no true claim wrongly called false. What it does not measure is the failure we care most about: a true statement with one shade of meaning moved, a consequence sold as a cause. We have not managed to write those test items yet. We would rather tell you that than tell you it is perfect.
How it is made, in plain terms
A page that demands sources should not be vague about itself, so: AI is part of how these videos are made. A model drafts the script and a person rewrites it. The voice is synthesised. The objects on the board are generated one at a time, then drawn by our own drawing engine, stroke after stroke, in the order the story needs them. That engine is ordinary software: no model decides where the marker goes.
What is not automated is the part you are trusting us with. Facts are checked by a different model from the one that wrote them, anything it disputes is decided by a person, and no script, board or claim list ships without someone reading it first. This page is where we say so, because a viewer should learn that from us rather than work it out.
We think that is the honest trade: machines for the drawing and the reading-aloud, people for the deciding. If that bothers you, the sources are on every page. Check us.
When we get it wrong
We will. When it happens: the correction goes in the description and on this site's page for that video, and if the error changes the answer, it goes out to The Redraw subscribers too. No silent edits, no quietly re-uploaded files.
Two things make that checkable rather than a promise. A claim our own check finds false gets the correction printed under the sentence that says it, not tucked in a table at the bottom where the piece can contradict itself in peace. And every one of them collects onone page, which builds itself from the checks, so it cannot quietly stop being updated.
Found something we got wrong? Write to hello@handorial.com. A correction with a source is the most useful email we can get.