How does this work?

You type a subject. A while later there is a deck of cards with facts on them, each one checked against a page somebody can go and read, and a picture we are allowed to print. This is what happens in between, including the parts we get wrong.

The short version

A language model does the research, in your browser, using the public catalogues of museums, libraries and archives — not a search engine, and not its own memory. Every fact it writes on a card has to be a sentence it can quote from a page we fetched ourselves. Every picture has to carry a licence that permits commercial printing, and a second model has to agree that the picture shows what the card says it shows.

Nothing on a card comes from the model's recollection. That is the whole design.

Where the facts come from

The deck decides what its cards carry — a lifespan, a nationality, a best-known work, a picture — and then each card is researched on its own. The researcher is a language model with seven tools: search an institution, ask for the detail of one record, open a page and read it, collect whatever has come back, record an answer, wish for a source we do not have, and finish.

It runs in your browser rather than on our server, which is not an implementation detail. Sixty-odd institutions each have their own rate limit, counted per client. A server asking on behalf of everybody exhausts an allowance that a thousand browsers asking for themselves never touch — and the Art Institute of Chicago refuses our server outright while serving your browser happily. So your machine talks to the archives, and ours holds the keys and does the checking.

It starts by reading, not by searching

The subject's encyclopedia article is opened and read before the model is asked anything at all. This sounds obvious and was not there for a long time; without it, runs spent their first three turns typing guesses into catalogue search boxes.

Every fact has to be quotable

When the model records an answer it must also supply the sentence it took the answer from, and the identifier of the page it read. Our server then checks that the sentence actually appears in the bytes we fetched. If it does not — and if no sentence in that page supports the answer — the answer is still kept, but it is marked unverified, and it says so in the editor.

A search result's snippet is never evidence. Only a page we opened and stored counts, because a claim you cannot go back and check is not a claim, it is a rumour.

Forty-four per cent of categories coming back empty is not a rounding error, and we would rather show you a gap than fill it with something plausible. A card that says nothing about a person's birthplace is a card whose sources did not say; a card that guesses is a card that lies in a way nobody can see.

Where the pictures come from

Eleven picture archives are asked: Wikimedia Commons, the Art Institute of Chicago, the Metropolitan Museum of Art, the Library of Congress, NASA, the Internet Archive, the V&A, the Wellcome Collection, the Cleveland Museum of Art, SMK in Copenhagen, and Openverse. Each has eight seconds to answer before the search goes on without it.

Then two things are checked, in this order, because they fail differently.

First: may we print it?

Every result is classified from whatever rights wording its institution supplies — the Met's rightsAndReproduction, Commons' UsageTerms, the Library of Congress's free-text rights field. None of them agree on a vocabulary and half of them write sentences, so the text is read rather than matched against a list of licence codes.

The rule underneath is simple and deliberately one-sided: permission has to be stated. Silence is not permission. An archive that returns a photograph with no rights field has told us nothing, and "nothing" is a refusal here.

Second: is it a picture of the thing?

A licence check does not know what is in a photograph. So the best few survivors are looked at by a vision model, which is asked whether the picture shows the subject itself — and is told to say no to the things a search cannot tell apart: a sign naming it, a museum about it, a replica of it, a portrait of somebody connected with it, a modern site where it used to be.

The catalogue entry travels with the picture, marked as a claim rather than a fact, because for a great many subjects it is the only evidence of identity there is: nobody can recognise a sixteenth-century warlord by sight, and neither can a model. We measured whether that helps or whether it just makes the model agree with the archive.

Which licences we print, and why

We sell these decks. That makes the question narrower than "is it a free licence": it is "may we reproduce this, commercially, on a physical object, and what do we owe in return".

The same classifier runs in your browser when the carousel is filled and again on our server when a picture is chosen, so a picture cannot be offered under one rule and stored under another.

Share-alike, and what it actually asks

This is the licence we refuse most often, so it deserves the whole argument rather than a verdict. An earlier version of this page said that one BY-SA photograph would mean "the deck must be published BY-SA". That was too strong, and it is worth correcting rather than quietly editing.

Share-alike cannot relicense anybody else's work, and we could not do it if it tried. You can only put a licence on something you own. A public-domain engraving does not stop being public domain because it shares a box with a BY-SA photograph; another photographer's CC BY picture stays theirs on their terms. What share-alike governs is the thing built from the licensed picture — the "adapted material" — and it asks that that be offered on the same terms.

So the real question is whether a card is an adaptation of the photograph on it, or a collection containing it. Creative Commons' own guidance is that putting an image in a book, a slideshow or a magazine page next to text is a collection: the image keeps its licence and the book does not catch it. Modifying the image itself — recolouring, compositing, tracing — is an adaptation. A card sits between the two: we crop the picture to its frame and set type beside it.

Our honest reading, and we are not lawyers: the photograph stays BY-SA and must carry its credit, its licence, a link and a note that it was cropped; the card face that carries it is at worst an adaptation, and would have to be offered under BY-SA; and the other cards, the other pictures, the text and the deck as a whole are a collection and are untouched. The exposure is one card, not fifty-four — and what it gives away is that card's design, not anybody's photograph.

We refuse it today because it is a decision for whoever is selling the deck, and because doing it properly means more than flipping a flag. Three things have to be true, and only the first is about printing.

• The credit, in print, next to the picture. We do this already.
• The licence named and linked, and the crop noted. A QR code and a short URL on the tuck box can carry that — the licence says the conditions may be satisfied "in any reasonable manner based on the medium", and a link to a page listing every picture, its photographer, its licence and its address is the ordinary way to do it on a physical object.
• Actually granting the licence. This is the one that a link does not do by itself. Share-alike is not a duty to make the work remixable somewhere; it is a duty to offer that card face on the same terms. A page that lets somebody build their own deck is a service, not a copyright licence — and if our terms of sale said the card designs were ours alone, that would be an additional condition on the adapted material, which share-alike forbids. So the page the tuck box points at has to say, in words, which card faces are published under CC BY-SA.

And one thing has to be watched for, which is why it is not simply a switch: licences that cannot share a card. A share-alike picture requires that whatever is built from it be share-alike; a stock photograph is licensed to us for printing and expressly not for passing on. Both are perfectly valid, and a card carrying both cannot be published at all — honouring one breaks the other. The same is true of a share-alike sentence quoted next to bought art. Every card is now read as a whole before a deck can be frozen for the printer, and a card whose licences contradict each other is refused by name, on the screen, while it is still an editing decision rather than a legal one.

The same question, about the words

Wikipedia's prose is CC BY-SA as well, and we read a great deal of it. The distinction that matters is the one copyright itself makes: a fact is nobody's property, the sentence somebody wrote to express it is theirs. "1773–1857" and "British aeronautics engineer" can be printed by anyone; copying a distinctive sentence onto a card would put that card under the same licence as the picture case above.

So the researcher quotes to prove an answer and states the answer in short form. The quotation is kept as evidence, shown to you in the card's research trace, and not printed on the card. We checked rather than assumed:

The picture on the card is not always the file

Archives hold a great deal of scholarship as TIFF, PDF and DjVu — the Metropolitan's arms-and-armour catalogue is a PDF, and a good deal of archaeology is scanned as TIFF. Nothing that draws a card can draw any of them. So when the file itself cannot be rendered we ask the institution for its own JPEG rendering, at the largest size it says it will actually serve, and print that. The licence and the credit are the file's; only the encoding changes.

What we still get wrong

• Categories come back empty. Roughly two in five, and more for subjects the free encyclopedias treat briefly. We show the gap rather than filling it.
• Some subjects have no printable picture at all. Not because the search failed — because every photograph of them is somebody's property. A modern building in a country without freedom of panorama is a common example.
• The classifier reads prose. An institution that invents a new way to phrase its rights will be read as "unknown", which means refused. We would rather lose a usable picture than print an unusable one.
• Unverified answers still appear, labelled as such, because an answer that no sentence supports is sometimes still the answer — and sometimes the model talking. The label is the honest part.

If you are checking our work

Every card's research is inspectable: what was searched, what was opened, what was quoted and which sentence it came from. Every picture carries its source page, its licence as the institution stated it, and the credit as printed. If you find a picture on one of our cards that we should not have printed, tell us and we will take it off and say why.