The New York Times calls AI’s founding act a theft
NYT publisher A.G. Sulzberger says the press has been too timid to fight AI. Now what?
A.G. Sulzberger does not often raise his voice. The chairman and publisher of the New York Times argues in the measured cadence of a man who expects to be quoted accurately. So when he took the stage for the opening keynote at the WAN-IFRA World News Media Congress in Marseille, where I was sitting in the auditorium, and accused the richest companies in history of an “original sin”, the phrasing was deliberate. The sin is theft: the appropriation of journalism, books, music and film to train artificial-intelligence systems, carried out, in his words, “without permission or compensation”, and at a scale no previous technology has attempted.
None of this is new from him. The New York Times has litigated a version of the argument against OpenAI, Microsoft and Perplexity for two and a half years, at a cost Sulzberger put above $20m. In Marseille the setting sharpened it. He had come to tell publishers from some 60 countries that their response to AI has been “too quiet, too passive, and too fragmented”, and that polite resignation now leads only to extinction. From where I sat, the hall heard him out in a sombre quiet. The New York Times is the great commercial success of the subscription era, with over 12 million paying subscribers, by far the largest number by some way, and a warning from the industry’s clearest winner is harder to dismiss.
What the engineers pay for, and what they take
Sulzberger’s primary argument was an accounting one. An AI model, he said, is built from four ingredients: the talent that designs the algorithms, the computing infrastructure of chips and data centres, the energy to run it, and what the industry blandly calls “data”. The first three are paid for, lavishly. No AI chief executive proposes that the best engineers work for nothing; pay packages run into the millions. None would steal chips from Nvidia or tap a power line. Investors are absorbing hundreds of billions in losses to build the plants.
The fourth ingredient is taken for free. And “data”, he noted, is a word that flatters the theft by making creative work sound like a commodity, lying around for collection. The justifications change with the occasion: innovation demands it, facts cannot be owned, licensing is too slow and too costly, fair use permits it, national security requires it lest China win the race. Each one, he argued, falls apart under scrutiny. A chatbot can recite facts only because it has absorbed the protected language and structure that carry them, and drafting a licence costs a great deal less than building a power plant or data centre. A country gains nothing in a technology contest by dismantling the intellectual-property regime that underwrites its own creative economy.
The six leading AI firms are worth some $11 trillion between them, by his reckoning, more than three times French GDP; private AI investment in the United States ran to roughly $350bn in 2025 and is climbing. Against that, the licensing deals struck with publishers’ amount, on the reported figures, to less than half of one per cent of the money flowing to the people who produced the training material. The money is plainly there; withholding it from creators is a choice, not a constraint.
The ingredient nobody wants to cost
If data is cheap, quality journalism is the part of it that AI firms most quietly prize. Sulzberger cited an admission from OpenAI that it would be “impossible” to train today’s leading models without copyrighted material, and an engineer’s blunter line that a model’s success is “determined by your data set, nothing else”: you are, in the old machine-learning saw, what you eat. Five of the ten most-used sources for training popular large language models are news publishers. In one major training set, he said, the Times was the single largest source of proprietary data, ahead of the Guardian and the Los Angeles Times. A Microsoft executive has conceded that premium content measurably improves a model’s answers.
That output is expensive to produce. Sulzberger said the Times published close to half a million pieces last year, in articles, photographs, video and audio, at a cost of more than $2bn, with correspondents in all 50 American states and 155 countries, more than 70 of its journalists in Ukraine. A chatbot values that material precisely because people reported, verified and edited it, people who can be sued, sacked or sent to a war zone. The AI industry has nonetheless maintained that it owes nothing for any of it. Meta was trained on a corpus of pirated books. Perplexity scraped sites that had explicitly refused it. OpenAI has lobbied Washington for legal immunity. Sulzberger reserved a pointed line for the firm with the cleanest reputation, telling the hall that “even Anthropic”, lauded for its ethical posture, had in his account been unwilling to pay for the journalism it uses.
Litigation is the obvious recourse and a poor one. The New York Times’s case has run for thirty months, and cost more than many newsrooms earn in a year. AI firms know that almost no one else can afford to follow it, which leaves the law to be tested by whoever has the deepest pockets.
Sulzberger makes the case as both plaintiff and dealmaker. The New York Times sued OpenAI and Microsoft in 2023 and has not settled; in 2025, it signed its first AI licence, with Amazon. OpenAI has separately struck paid deals with other publishers, among them News Corp and Axel Springer. The pattern complicates his arithmetic and confirms his warning at once: the marquee titles are being paid something, while the smaller newsrooms he was addressing are offered nothing. His quarrel is less that money never changes hands than that it does so on the buyer’s terms, for a fraction of the value, and after the work has already been taken. There is also a circularity to his position. Sulzberger calls the practice theft; but whether it is theft or fair use is precisely the question his own lawsuit exists to settle, and no court has ruled yet. He is, in effect, asserting the verdict while the trial is still running.
The click that no longer comes
The deeper threat lies in substitution rather than training. For two decades, the open web ran on a lopsided bargain: search and social platforms took a growing share of advertising but sent readers back to publishers, who could sell subscriptions or show ads against the traffic. Google set the terms and kept sending the clicks.
That exchange is being withdrawn. As Google answers more queries directly, the link that used to deliver a reader is, by his estimates, roughly 10 times harder to earn than it was a decade ago. Rival AI products are stingier still: one study put their referral rate at 96% below a Google search. The largest newspapers tracked by Comscore have lost more than 45% of their traffic on average over four years, with the slide expected to continue. Sulzberger quoted Microsoft’s own head of AI monetisation, conceding that the old value exchange “does not translate cleanly to an AI-first world”, then read out the launch copy for its AI search, which invites users to skip the links and “talk through whatever you’re curious about”: an honesty about the model that its own commercial behaviour does not extend to publishers.
The revenue picture behind all this is stark. Newspaper advertising income has fallen by around 80% over twenty years, according to his figures, while Meta alone now earns eight times what all newspapers on earth make from ads combined. Subscriptions were the lifeboat. A reader who can extract the same reporting free from a chatbot has little reason to pay for it, and around a third of AI scrapes, he said, breach explicit restrictions, paywalls among them. Licensing cheques and per-scrape micropayments will not, for most publishers, replace what is lost; the majority expect no meaningful AI income at all.
Confidently wrong
The case becomes harder to wave away when the products fail at the one task they claim to perform. Sulzberger cited European Broadcasting Union research findings that leading AI assistants misrepresented the news in close to half of all answers. When the activist Charlie Kirk was killed last year, Perplexity’s bot suggested the White House statement was fabricated and Grok insisted he was alive. Because the systems handle uncertainty poorly, they tend to be wrong with conviction and neither track nor correct their errors. Microsoft, he noted, had launched Copilot with a disclaimer telling users to rely on it at their own risk.
Then he turned to what comes next. Amazon Web Services reckons most online content is already machine-generated, perhaps nine in ten before long. Fake local-news sites already outnumber real ones, cheap to erect as the genuine article grows harder to sustain. Sulzberger borrowed Margaret Atwood’s image of the author “murdered by my replica”. The hollowing of original reporting starves publishers and corrodes the shared factual record on which a society agrees to disagree, and on which the chatbots themselves depend.
The list of demands
His prescriptions were practical and unusually concrete. Defend your copyright and litigate when you must. License, but only on terms that reflect sustainable value and leave you some say over how the work is used. Lobby legislators with a short common list: bots that identify themselves, transparency over how content is used, liability for defamatory output, copyright protections strengthened not weakened. Above all, act together, and with the music, film and book industries facing the same theft, rather than cutting deals one frightened newsroom at a time.
The second half of his advice turned inward. Adopt AI inside the newsroom aggressively and under human control, because the laggards will be flattened anyway. Build a direct relationship with readers, so the platforms cannot ration the audience. Recommit to original reporting, the investigations and eyewitness accounts, because it is the one product a chatbot cannot manufacture and the public cannot get elsewhere.
In the last great disruption, Sulzberger admitted, The New York Times swallowed Silicon Valley’s line that “information wants to be free”, forgetting that Stewart Brand’s original aphorism had a second clause: information also wants to be expensive, because it is valuable. He asked the room not to be naive twice.
He stopped short of an answer, because it is not his alone to give: either the value of journalism returns to the people who go out and find it, or it stays with the firms that take it. The wave already breaking, he warned, was the small one. The New York Times, with its lawyers and its licensing deals, will probably ride it out. Whether the rest of us in the hall can do the same was the question he left us holding.