Who Gets Paid When the Machine Answers: The New Economics of AI Licensing

The lawsuit that named the terms

For most of the web's history, a publisher's business model rested on a simple exchange. A search engine sent a reader to a page, and the page carried an advertisement that paid for the visit. An AI answer engine changes that exchange at the root: it reads the page, folds a sentence from it into a generated answer, and the reader who receives that answer frequently never visits the original page at all. The click that used to fund the article is no longer guaranteed to happen.

The dispute over what that change owes the publisher reached a court on December 27, 2023, when The New York Times filed suit against OpenAI and Microsoft in the Southern District of New York. The complaint alleged that millions of the paper's copyrighted articles had been used to train ChatGPT without consent and without payment, and it named Microsoft as a defendant alongside OpenAI because of the financial and technical partnership between the two companies. It was among the first major publisher lawsuits of its kind, and it set the terms the rest of the industry would argue over for the next two years: was training a language model on a copyrighted article a research use, protected the way earlier courts protected search engines that indexed the same page, or was it a use that required a license and a payment, the way republishing the article itself would?

OpenAI answered in public. In a blog post on January 8, 2024, the company called the Times' claims 'without merit,' arguing that training a model on text already available on the open internet fell within the fair use doctrine that has long let search engines, archives, and researchers work with copyrighted material without paying for each use. That argument, that mass ingestion of public text for machine learning is a research use rather than a competing republication, remains the central legal position of every AI lab now facing a similar suit. Whether it holds is not yet settled, and the uncertainty is exactly what pushed a second track to open alongside litigation: instead of arguing the point in court, a publisher could simply get paid.

The price of being licensed

OpenAI answered that opening with money as often as with argument. In 2024 the company signed a licensing deal with News Corp, Rupert Murdoch's publishing group, worth more than $250 million over five years, one of the largest agreements disclosed in the new market. The deal covers content from The Wall Street Journal, the New York Post, and the Daily Telegraph in the UK, and it lets OpenAI train on and cite that reporting inside ChatGPT under contract rather than under contest. For a publishing group that owns some of the highest-authority mastheads in English-language business and political reporting, the deal converted a legal exposure into a revenue line.

Reddit found a similar opening from the other direction, licensing the same underlying asset, its archive of user discussion, to two buyers at once. In early 2024 the company signed a deal with Google reported at roughly $60 million a year to train Gemini, and separately reached an agreement with OpenAI reported at roughly $70 million a year, a combined figure close to $130 million a year flowing to one company from two AI labs. Reddit's position was structural: its forum threads capture the kind of first-person, conversational text that formal news writing rarely produces, and that a chat-style model needs in order to sound like a person rather than a press release.

A third deal, reported in 2026, showed the same publisher can keep collecting. Meta signed its own AI content-licensing agreement with News Corp, valued at roughly $50 million, a figure still emerging in the reporting rather than independently confirmed at the level of detail attached to the OpenAI deal. Stacked against the News Corp-OpenAI number, it suggests a publisher with a large enough archive and enough legal standing can sell training rights to more than one buyer without exhausting the asset, the way a single piece of music can be licensed to several films.

Across the market as a whole, disclosed deal values have ranged from roughly $5 million to $250 million per agreement between 2023 and 2026, according to a market tracker published by Quartz. That is not a price for a unit of content; it is closer to a price for a relationship, set by a publisher's archive depth, its legal standing, and how badly a given AI lab wants its brand associated with a credible newsroom rather than an anonymous scrape. No visible market rate has formed. Each deal is negotiated from a blank page.

The publishers who chose to sue instead

Not every publisher got a call from an AI lab's business-development team, and not every publisher that did took the deal. News Corp's own Dow Jones and New York Post sued Perplexity AI on October 21, 2024, alleging what the complaint called 'massive' and 'illegal' copying of copyrighted articles by the AI search company. The suit was amended on December 11, 2024, to add claims of trademark dilution, arguing Perplexity's habit of attributing fabricated or altered content to Dow Jones and Post bylines damaged the mastheads' names directly, not just their revenue.

Reddit, one of the best-paid publishers in the new licensing market, sued a different AI lab entirely in June 2025. The complaint against Anthropic alleged the company scraped and used Reddit's user data to train its models without ever signing a license, the same conduct Reddit had already negotiated a paid agreement to permit for Google and OpenAI. The suit is a clean natural experiment: the same content, the same publisher, one buyer that paid and another the publisher says did not, and only one of the two ended up in court.

What links the Perplexity and Anthropic suits is not the publisher; it is the absence of a deal. A company that scrapes without an agreement inherits exactly the legal exposure OpenAI tried to defend against in the New York Times case, and inherits it from every publisher with the resources to notice the scraping and file. The publishers without those resources, the great majority of the web's writers, sit in a third category this reporting does not directly measure: neither paid nor suing, because most have neither the archive depth nor the legal budget to be worth negotiating with or worth defending against in court.

The same masthead, two strategies

News Corp's own record makes the cleanest case that this is not a divide between publishers who cooperate and publishers who resist. The same company signed licensing deals with OpenAI and Meta while its Dow Jones and New York Post units sued Perplexity, all inside the same eighteen-month window. The company did not choose a single posture toward AI companies; it negotiated a separate answer for every buyer, license where a buyer offered a fair price, sue where a buyer did not offer one at all.

An industry tracker following the licensing market through 2026 describes the wider pattern the same way: publishers have split into a license camp and a sue camp, but the split runs by relationship rather than by company. A publisher licenses the buyer willing to pay and sues, or waits to sue, the buyer that is not. The strategy variable is not the publisher's philosophy about AI; it is whether the specific company on the other side of the negotiation chose to make an offer.

The same publisher takes both positions inside a single year, which rules out reading this as a referendum on whether training AI on news content is acceptable. What is actually happening is a negotiation, buyer by buyer, over whether a company with the scale to train a frontier model would rather pay a licensing fee or risk a multi-year lawsuit and an uncertain fair-use ruling. So far, the largest labs have often chosen to pay at least some publishers, which is itself a signal about how seriously they read their own legal exposure.

A four-company buyer's market

Every deal recorded here sits on one side of the table with a strikingly small set of names. OpenAI, Google, Microsoft, as OpenAI's primary financial partner and co-defendant, and Anthropic account for essentially the entire disclosed market for AI content licensing and for the lawsuits that stand in for it. All four are headquartered in the United States. The sellers on the other side of these deals include a British-Australian-American media conglomerate, an American internet forum, and, in the lawsuits, the same set of publishers plus whichever outlet a given AI company scraped without asking.

That asymmetry is the geopolitical shape of the market, not an incidental fact about it. A newspaper in India, a broadcaster in Brazil, or a wire service in continental Europe has no domestic AI lab of comparable scale to negotiate with instead. Its practical choices narrow to the same three this reporting has traced inside the US market: license to one of four American buyers, sue one of the same four American buyers in an American court, or absorb the scraping with no forum and no payment at all. The News Corp deal already reaches across that boundary once, since it folds in the Daily Telegraph, a British masthead, inside a contract negotiated and priced in the United States.

Every dominant medium in this history has done both things at once: it has opened a new channel for a shortlist of participants and concentrated authority over everyone else. The licensing deals genuinely open a new channel for the publishers who land one, handing News Corp and Reddit a real, disclosed, non-advertising revenue line that did not exist three years ago. The same deals concentrate the terms of a global content trade, price, scope, and payment schedule, inside four boardrooms that set each contract unilaterally, publisher by publisher, with no visible common rate and no publisher-side coalition large enough to negotiate back.

What the courts have not settled

The legal question underneath every one of these deals is still open. A federal judge in the Southern District of New York ruled on April 4, 2025, allowing the core copyright claims in the Times' suit against OpenAI to proceed, rejecting OpenAI's motion to dismiss on several counts. That ruling is procedural. It means the case survived an early attempt to end it, not that a court has decided whether training a model on copyrighted text without a license is fair use. The Reddit suit against Anthropic, filed two months later, raises the identical unresolved question from a different publisher's chair.

Until one of those cases, or another like it, reaches a final verdict, the licensing deals already signed function as a private answer to a question the public court system has not yet given. News Corp, Reddit, and the other publishers with disclosed deals in the $5 million to $250 million range are effectively setting a market price for AI training rights ahead of any settled law, one contract at a time, while every publisher without a comparable deal waits on a ruling that could still go either way.

The dispute traced here sits one step upstream of a second question this publication tracks directly: when an answer engine names a source inside a generated answer, the reader sees the name, not whether the source behind it was ever paid to be there. The lawsuits and licensing deals in this record settle, or attempt to settle, who gets compensated for supplying the material a model trains on. They do not settle who a model chooses to cite once it is answering a live question, a related decision the industry has so far negotiated in public far less than it has negotiated the first.