The Provenance Question Went to Court
Watch on YouTube · subscribe for the daily brief
Two days ago a mathematician accused OpenAI of racing to publish a proof that leaned on his leaked work. Today authors told a federal judge OpenAI built ChatGPT itself on pirated books, torrented from a shadow library, then hid the evidence by renaming the datasets. The provenance question is no longer a debate. It's a lawsuit with internal documents attached.
You found the story again, and it is the right one. On the surface it is another AI copyright lawsuit, and there have been many. But this filing matters more, for two reasons. First, it is specific and it is armed: internal documents, a summary-judgment motion, a direct accusation of concealment. Second, and this is what makes it the story of the week rather than just another docket entry, it lands two days after the Navier-Stokes scandal, where a mathematician accused OpenAI's models of leaning on his leaked work. Same question, different room. Where did the intelligence actually come from, and was any of it taken? On Tuesday that was an ethics dispute. Today it is a federal case with the receipts allegedly attached.
Here is what happened. A proposed class of book authors filed a motion for summary judgment asking US District Judge Sidney Stein to rule, before any trial, that OpenAI copied their work without permission and that it does not qualify as fair use. The language is not lawyerly hedging. The filing states: "It is no exaggeration to say that OpenAI built the foundations of its business on mass piracy." It alleges OpenAI obtained books not by buying them but by torrenting them from LibGen, a notorious shadow library, and then hid the evidence by renaming the datasets. The brief covers 194 titles, seeks a finding of liability rather than damages at this stage, and frames the stakes bluntly: "OpenAI's GPT models pose an existential threat to those who write and publish books," while "AI-generated books of all types are already flooding the market." OpenAI filed the mirror-image motion the same day, arguing its data harvesting is fair use, that its models almost never reproduce the articles they read, and that the plaintiffs have grown, not shrunk, since ChatGPT launched. Much of both filings is redacted, black bars over the traffic figures, the revenue comparisons, the internal executive commentary. So the public is seeing the shape of the fight, not its full contents.
The legal detail that makes this genuinely dangerous for OpenAI, and worth understanding even if you never read a docket, is a shift in the authors' theory. Early in the AI copyright wars, plaintiffs tied the piracy to the training as one act: you stole the books and you used them to build a competing product. Now they have separated the two. The new argument is that the illegal download itself is copyright infringement, independent of whether the book was ever used to train anything. That is a much harder charge to escape, because OpenAI cannot defend it by pointing to how transformative the training was. And there is precedent that should worry them: in the Anthropic case, a judge wrote that "that Anthropic later bought a copy of a book it earlier stole off the internet will not absolve it of liability for the theft." The act of pirating stands on its own. If Judge Stein applies that logic here, the "but the model is transformative, so it's fair use" defense does not even reach the torrenting. OpenAI could be liable for the download regardless of what it did next. That is the theory that already drew blood against Anthropic, and OpenAI is standing in front of it now.
Why this matters for you, the person building on these models, is the uncomfortable part, and it is the through-line from Navier-Stokes to here. Every model you deploy was trained on something, and this week a court began deciding whether that "something" was lawfully obtained, and what happens if it wasn't. There is a real, if still distant, scenario where a foundational model is ruled to have been built on infringing data, and that ruling creates legal exposure that flows downstream to the products built on it. We are not there. But the direction is set, and the smart move is not panic, it is provenance awareness. The same instinct that should make you cautious about what you type into a vendor's tool (the Navier-Stokes lesson) should make you ask what your vendors can actually tell you about their training data, and what indemnification they offer if it turns out to be tainted. The major providers have started offering copyright indemnities precisely because this question is coming for their customers, not just for them. Read those terms. This is no longer abstract; it is a discovery fight with 20 million ChatGPT logs already ordered produced and internal dataset-renaming allegations on the record.
Now the honest counterweight, because skepticism cuts both ways and this is a motion, not a verdict. These are the authors' allegations, filed to persuade a judge, written in the strongest possible language, and heavily redacted so we cannot fully evaluate them. OpenAI's fair-use argument is not frivolous: a California judge handed Meta a "bittersweet victory" on similar claims, and the law here is genuinely unsettled. OpenAI's economic argument, that the Times and other plaintiffs have grown revenue since ChatGPT, is real, the Times reported Q2 digital ad revenue up almost 21 percent. Fair use is decided case by case, and no court has yet ruled that training a large language model on copyrighted text is categorically infringement. So the accurate status is: serious, specific, document-backed allegations that the acquisition of the data was illegal, meeting a genuine and unsettled fair-use defense, in front of a judge who has already ruled against OpenAI on discovery. Not a conviction. But not noise either. The provenance question, which this whole run has circled from the semantic layer to the Navier-Stokes fight, is now being formally adjudicated, and how Stein rules will shape what every model built on scraped data is worth.
The rest of the day was a clean builder's bench, and it fits, because most of it is about controlling and understanding your own data rather than borrowing someone else's. Starburst's "What is Column-Level Security?" is genuinely useful fundamentals, controlling access at the column level is exactly the governance discipline that matters more as agents query your warehouse. Databricks' "Evaluation-First AI Agents," on how Zepto scales customer support with MLflow, is the standout: evaluation-first is the recurring right answer for production agents, and a real customer case is worth more than a framework. Supabase landed in Gemini Enterprise, and Vercel made production-deployment protection free on every plan while moving password protection to per-project pricing on Pro, small but real if you run on Vercel. On the voices side, three strong technical reads: Sebastian Raschka's "GPT-6 Astra, Looped Transformers, and Hidden Reasoning" is the deep dive on what actually changed under Astra's hood, and it pairs with the chain-of-thought worry from Pachocki's essay; Damien Benveniste's speculative-decoding explainer is a clear piece on making models faster; and Ben Lorica's "Self-Improvement Without the Science Fiction" is the grounded take on the recursive-improvement talk that OpenAI's 10,000-agent run stirred up two days ago. And Simon Willison quoted Terence Tao, almost certainly on the Navier-Stokes work, a nice thread tying today's provenance-in-court story back to Tuesday's provenance-in-math one.
The through-line, one sentence: on September 10 book authors asked a federal judge to rule that OpenAI built ChatGPT on mass piracy, torrenting books from a shadow library and hiding it by renaming datasets, and the reason it matters beyond one lawsuit is that it is the same question the Navier-Stokes scandal raised two days earlier, where did the intelligence come from and was it taken, now moved from ethics into a courtroom with internal documents, a dangerous new legal theory, and a judge who has already ruled against OpenAI once.
One thing to do today: turn provenance from a headline into a checklist item for your own stack. You do not control how the foundation models you use were trained, but you can control your exposure to the question. Two concrete moves. First, read the copyright indemnification terms your model vendor offers, actually read them, because the major providers added those clauses specifically because this litigation exists, and knowing what you are and are not covered for is now part of responsible vendor selection. Second, apply the same discipline to your own data pipeline that this case is about: know the provenance of the data you train on or fine-tune with, because "we scraped it and it was probably fine" is precisely the defense being tested in front of Judge Stein right now, and it is not going well. The lesson of this week, from a math proof to a courtroom, is the same one this whole run keeps landing on: the value and the risk both live in where the data came from. Know where yours came from.
Sources covered
1 story · 10 sources (0 news, 5 company, 5 voices)