How to price your journalism for AI when you're flying blind

Since last September there has been a way to put a price on your journalism that machines can read. It is called Really Simple Licensing, or RSL, and it works like a rate card sitting on your server: Training on an article costs one thing, using it to answer someone's question costs another, quoting a passage costs something else again. Any AI crawler that arrives can read the card before it takes anything. 

Reddit, Yahoo, Medium, People Inc. and O'Reilly are behind it, a WordPress plugin has made it a five-minute install since July and declaring your terms costs a publisher nothing. Worth doing, because that file does a second job: it is also a "no" written so a machine can act on it, which is what EU law asks for and what a German court in December found ordinary website wording may not deliver. What it can't tell you is who read it because the agents still identify themselves on the honor system. 

For thirty years, publishers had one instrument for talking to crawlers: robots.txt, a file that says come in or stay out. It cannot say "you may quote me but not train on me," and it cannot mention money.

RSL can. Training, answering a question, caching, quoting, reproducing a whole article are separate permissions, each with its own price, written in a file the crawler finds on arrival. It launched in September 2025 and is run by a nonprofit, the RSL Collective, whose model comes from music: Pool the rights of publishers too small to negotiate alone, the way performing-rights societies do for songwriters.

Cosmin Ene, the chief executive of Supertab, which sets up RSL for publishers, is careful about what this achieves. A crawler that reads the file and takes the content anyway, he says, is bound by terms it has already been shown: "If you don't take that warning seriously and keep using my content, then the terms I communicated to you apply," he told me. But those terms aren’t enforceable yet. RSL is how you put the claim on record. Making it stick is a separate fight, in court, with different answers in different countries.

Here is what the rate card runs into. When a crawler requests a page, it announces who it is in a line of text that it writes itself. But nothing checks that line. Anyone anywhere can send a request that says “ClaudeBot” and the server receiving it has no way to know better.

There will be a fix. Instead of taking the name on trust, a crawler can sign each request with a key only its operator holds, so the publisher checks the signature rather than believing the label. Cloudflare leads that work, with Amazon, Akamai and OpenAI behind it, and Supertab says the version of its product shipping in September will look for those signatures where they exist. The gap is on the other side: Hardly any crawler signs. Google says so in its own crawler documentation — not all of its user agents use the protocol.

European publishers are in a different position from American ones. The EU's 2019 copyright directive, written into German law as §44b UrhG, lets rightsholders refuse text and data mining, which is the legal category AI training falls into. There is one condition: The refusal must be made in a form machines can process.

What that means in practice was tested in December, when the Hamburg appeals court ruled in the case of Kneschke v. LAION. The reservation a photographer relied on had been written in ordinary language on a stock agency's website, and it barred "automatic evaluation" rather than text and data mining, so it had to be interpreted before anyone could act on it. The photographer lost. The court would not assume that wording meant for human readers counted as machine-readable in 2021. The court left open whether it counts now, when models read prose perfectly well. The case is on appeal to the Federal Court of Justice.

German rights holder organizations read the decision as a setback, and for that case it was. The lesson for a publisher points the other way. A refusal a machine can read without having to work out what it means avoids the weakness the court identified. A license file is exactly that: set choices, nothing left to interpretation. It costs a German or Austrian publisher little effort, and nobody can later argue they couldn’t see it.

The prices in the file are a different matter. Nobody has to accept them. What European law backs is the “no,” not the price.

The European Commission will soon decide which machine-readable formats general-purpose AI providers will be obliged to honor under the AI Act, working from a study by the EU Intellectual Property Office. Whether RSL ends up on that list matters more to a publisher in Hamburg or Vienna than which vendor sells the installation. 

Outside the EU there is no equivalent. An American publisher's file is a notice, and what it is worth rests on arguments no court has tested.

The list of backers is long: Reddit, People Inc., Yahoo, Ziff Davis, wikiHow, Medium, The Daily Beast, Raptive, O'Reilly and others, with Fastly, Quora and Adweek endorsing the standard without joining.

Endorsing, however, is not installing. Asked whether any AI company honors an RSL file today, Ene said there are barely any integrations yet. Supertab's own marketing claims more than 1,500 publishers have endorsed licensing standards like RSL, a figure with no published methodology behind it.

So, the standard is open, several companies will set it up for you, and almost nobody has switched it on. That is the current state. But it’s not the verdict. Adoption like this moves when something forces it, and the things that could force it – a court decision, a regulator's list of approved formats, a first invoice that actually gets paid – are all closer than the installation numbers suggest.

Ene's argument for why this matters is also his sales pitch, and the two are worth separating. Big AI companies will keep signing deals with big publishers, he says, but that model cannot stretch to tens of thousands of newsrooms. The internet works because systems do not renegotiate their plumbing every time they talk to each other, and he expects licensing to end up the same way: agreed formats, infrastructure both sides can find each other through, and payment that happens without a phone call. Supertab's position, in his words, is not to be the marketplace but the rails a market runs on.

He also thinks local publishers underrate the value they hold. A national paper can tell an AI system a great deal about France. If someone asks what happened this morning in one neighborhood in Marseille, that may exist only with the outlet that reported it, and no national licensing deal reaches it. He does not promise this turns into money soon – the economics are still forming – but he argues local newsrooms should be measuring and setting conditions now, so they know what they have before anyone offers to pay for it.

Declaring your terms costs little effort. Knowing whether anyone honored them is the expensive part, and it is the part a small newsroom cannot do alone. A large publisher has engineers who can read server logs, and probably pays a service like Cloudflare or Akamai that already sorts incoming traffic by who or what it is. A ten-person newsroom has neither, so it pays a vendor, and what arrives is a label on a dashboard: This visitor was a training crawler, that one was an assistant fetching a page for a reader. 

You cannot verify those labels yourself, and the vendor's answer about how it produces them tells you a lot about what you are buying. But the exercise still pays off before any of the licensing does. A newsroom that knows how much of its traffic is machines, which ones keep coming back and what they go for, understands its own distribution better than one waiting for a market to appear. That knowledge is useful whether anyone ever pays for the content.

Earlier in this series: whether to license at all, what to expose to agents, how they might pay, and how you would know what they took and whether your visibility is worth anything.