The AI Brief

U.S. District Judge Araceli Martínez-Olguín granted final approval Monday to Anthropic's $1.5 billion class-action settlement with authors and publishers, the largest known settlement in the history of US copyright law and the first major AI training lawsuit to fully resolve. The case, Bartz v. Anthropic, was filed in 2024 by authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson, who alleged Anthropic downloaded more than seven million works from pirate libraries including Library Genesis and Pirate Library Mirror to assemble a training corpus.

The governing ruling, issued by now-retired Judge William Alsup in June 2025, split the liability cleanly: training Claude on lawfully acquired or purchased books constituted fair use, Alsup described generative AI as "quintessentially transformative", but maintaining a digital library of pirated copies was independently unlawful regardless of how many were used in training. A damages trial had been scheduled for December 2025, with potential statutory liability estimated in the hundreds of billions; Anthropic settled to avert that exposure. The settlement pays approximately $3,000 per work across an estimated 500,000 works. Anthropic deputy general counsel Aparna Sridhar noted that more than 91% of eligible authors have claimed their share.

Judge Martínez-Olguín rejected objections from dissenting authors who argued the settlement was too small, ruling those complaints were "not grounded in a realistic assessment of the overall risks and rewards of a trial." She awarded $101 million of the $187.5 million in requested attorney fees. Some authors opted out of the class and continue to pursue separate lawsuits against Anthropic. Because Anthropic settled rather than appealing Alsup's fair use ruling, that ruling remains binding only at the district level, it does not bind other circuits or create controlling appellate precedent, meaning every other AI copyright case proceeds on its own record.

Disclosure: Claude, which generates this brief, is built by Anthropic.

The July 17 general-availability target for Gemini 3.5 Pro passed without a public launch, per reporting from Mashable and FindSkill.ai citing a Bloomberg report from July 16. Google has now missed the model's June GA date (announced at Google I/O on May 19), a subsequent July window, and the July 17 internal target, the fifth consecutive miss since CEO Sundar Pichai told a developer audience at I/O, "give us until next month." As of July 21 the public Gemini API shows no gemini-3.5-pro model ID.

The underlying cause, per Bloomberg and TechTimes reporting, is a full base-model rebuild after engineers found structural failures in recursive tool-calling and multi-layered SVG generation, failures the prior architecture could not close. The rebuilt model reportedly introduces a 2-million-token context window and a Deep Think reasoning layer. Google provided a statement to Mashable: "We're shipping quickly across a wide range of models while keeping them highly cost-effective for customers. We're currently testing 3.5 Pro, an upgraded Flash model, and other models with partners, and we're productively engaged with the U.S. government on model testing and broader frameworks." Prediction markets have moved the expected window to late July or early August, Polymarket pricing July 31 at approximately 81%.

Judge Alsup's June 2025 ruling in Bartz v. Anthropic, now final and paying out, establishes two distinct legal acts: (1) training a model on lawfully acquired books, which is fair use because generative AI is "quintessentially transformative" and does not displace demand for the original works; and (2) maintaining a central library of pirated books, which is independently infringing regardless of how many were actually used in training. The liability attaches to possession and use of unlawfully obtained copies, not to the training itself.

The parallel Ninth Circuit ruling in Kadrey v. Meta reached a similar fair use outcome but declined to separately analyze the lawfulness of how Meta's data was obtained, an analytical divergence that leaves the piracy question unresolved in that case. Because Anthropic settled rather than appealing, Alsup's ruling never reaches the Ninth Circuit and creates no binding appellate precedent. Each new AI copyright case must develop its own record. The New York Times' suit against OpenAI and Microsoft in the Southern District of New York, now in consolidated discovery including a court-ordered production of roughly 20 million de-identified ChatGPT logs, will be the next major test, and it is proceeding on different output-substitution facts.

A structural compliance gap remains for developers deploying globally: the Bartz fair use ruling applies only in US courts. Under Article 53 of the EU AI Act and the DSM Directive's text-and-data-mining provisions, GPAI providers must honor rights-holder opt-outs from EU-based training data regardless of fair use arguments in the US. US companies training under fair use assumptions and deploying models in EU markets must separately satisfy EU opt-out compliance, and the EU's Court of Justice is not bound by Alsup's analysis.

Gemini 3.5 Pro has now missed five consecutive public launch commitments, the original June general-availability window announced by Sundar Pichai at Google I/O on May 19, a subsequent July window, the widely-reported July 17 internal date, and a broader July window, and remains absent from Google's public API model list as of July 21. Bloomberg reported on July 16 that the model fell short of Google's internal quality goals on hallucination rates and real-world reliability after a full base-model rebuild, triggering a fifth delay. Engineers, AI researchers, and managers inside Google have expressed frustration that the company risks losing market position to Anthropic and OpenAI, per Bloomberg.

Google's public statement to Mashable acknowledges concurrent testing of "3.5 Pro, an upgraded Flash model, and other models with partners." That "upgraded Flash model" reference aligns with earlier reporting that Google has registered Gemini 3.6 Flash as a potential interim release, a stopgap tier that would arrive before 3.5 Pro if the rebuild continues to slip. Prediction markets price the next window at July 31 (~81% on Polymarket) and August 7 (~73%). First covered in Vol. I, No. 54 (July 18, 2026).

The EU AI Act's GPAI obligations under Article 53 took effect August 2, 2025, and apply to every provider placing a general-purpose AI model on the EU market, including US labs whose models are API-accessible from EU addresses. The DSM Directive's Article 4 text-and-data-mining exception permits rights holders to opt out of having their works used for commercial AI training; Article 53 of the AI Act makes respecting those opt-outs a legal requirement for any GPAI provider, backed by the AI Office's enforcement authority.

The Bartz ruling, that training on lawfully acquired US books is fair use under US copyright law, has no force in EU courts and does not satisfy Article 53. A US lab that legally acquired training data and trained its model in the US may nonetheless face EU compliance exposure if any data source was EU-origin, if the rights holder had filed a valid DSM opt-out, or if the lab's training-data transparency summary is insufficient. The European Parliament's March 2026 non-binding resolution proposed a 5–7% of global-turnover licensing fee as a stronger mechanism, signaling political momentum toward tighter requirements in the 2026 Copyright Directive review.

The practical compliance surface for AI developers is: (1) a copyright compliance policy posted publicly; (2) a training-data summary document accessible to the AI Office; and (3) an opt-out log showing which rights-holder requests were honored. No US-based frontier lab has yet published a complete Article 53 training-data summary. The AI Office is in its first enforcement cycle, with formal guidance on transparency summary standards expected by Q4 2026.