Docket Shock: AI Training Data on Trial

• Learn how two June 2025 rulings pushed U.S. courts to call large-scale Al training “transformative” fair use-while still leaving major liabilities on the table.
• Explore the major IP disputes poised to shape our evolving Al
• Gain insights from practice tips and the big questions surrounding Al and IP.

Artificial intelligence (Al) copyright fights are coming to a boil this summer. June alone saw two rulings that nudge the fair use line in the Northern District of California.

In Bartz v. Anthropic, the court held that using copyrighted books to train an (LLM) can be “exceedingly transformative,” qualifying as fair use. The court accepted Anthropic’s argument that they copied plaintiff’s books to “study” the style and themes, and “extract uncopyrightable facts” to build new technology. But there is a catch–the court declined to bless the company’s storage of a searchable trove of pirated e-books, ordering a separate trial on how those copies were obtained and whether statutory damages apply. Training on books without permission may be transformative enough to qualify as fair use for Anthropic, but the scraping and storage of ill-gotten content could cost them dearly-­ especially if class action status is granted.

Just days later, a different court in the same district granted yet another Al victory in Kadrey v. Meta (N.D. Cal., June 2025) – ruling that using books from several prominent authors to train Llama 2 was “highly transformative,” and could qualify as fair use. But the opinion reads like a road map for how to improve pleadings–the court criticizes that plaintiff’s claims failed to show reproduction of protected material or market harm.

And there are dozens of Al-copyright suits still on deck! Whether these June rulings prove to be a precedent or just a prelude, the Al playbook for copyright is being penned as we speak, and counselors need to keep up.

Fair Use Framework

Both June rulings rest on the same conclusion: training an LLM on entire books can qualify as fair use.

Fair use is the doctrine that permits limited use of copyrighted material without permission under certain circumstances, primarily for purposes like commentary, criticism, education or research. Courts assess fair use based on four factors.

1. Purpose and character of the use. Is the use transformative? Does it add new expression or meaning-or is it merely a reproduction of the original?
2. Nature of the copyrighted work. Creative works are afforded stronger protection than factual works.
3. Amount and substantiality of the portion used. How much of the original work was used, and was it the “heart” of the material?
4. Effect on the market. Does the use harm the market value of the original work or its potential licensing opportunities?

In both June rulings, the same conclusion was reached: copying entire books to train a LLM is “transformative” enough to qualify as fair use by the defendant. In both opinions, the courts emphasized that repurposing text as statistical training data serves a fundamentally different function from reading, licensing or selling the original works. That transformative leap can outweigh even wholesale, commercial copying.

But this decision framework does not close the door to plaintiffs entirely. Both opinions highlighted clear vulnerabilities. In the Anthropic matter, the court separated fair use from the issue of how Anthropic obtained its training set, refusing to dismiss claims tied to the company’s “searchable trove” of books scraped from shadow libraries, and ordering a separate trial on whether ingesting and storing pirated copies creates independent liability, statutory-damages exposure, and potentially even class certification. So, while fair use may cover the use of books as pre-training data, it doesn’t appear to launder illegal downloads or justify piracy.

The Meta ruling exposed a different crack: a call for more specificity. The court faulted plaintiffs pleadings for a lack of side-by-side passages showing that Llama 2, in fact, reproduces protectable chunks of text. And when it came to the most crucial fair use factor, market harm, the court described plaintiff’s arguments as “half-hearted,” elaborating that they chose to “put forward two flawed theories of market harm while failing to present meaningful evidence on the effect of training LLMs like Llama with their books on the market for those books.” The court invited an amended complaint with precise prompts, matching snippets and a showing of harm to the market value of the original work.

Whether these rulings become bedrock precedent or mere opening moves remains to be seen. Appeals are inevitable, parallel suits are stacking up in other circuits, and regulators are responding to calls for standardized guidance. It may still take years for the Al and fair use dust to settle.

Who to Watch: The AI/IP Matter Matrix

The fireworks have just begun. Across the country, disputes over training data continue to raise pressing questions. Here are a few of the major intellectual property (IP) disputes poised to shape our Al landscape.

These matters will continue to probe both fair-use factors and adjacent questions. Anthropic’s pirated-e-book situation may foreshadow similar scrutiny of “shadow-library” or torrent-sourced data in Getty, New York Times and Universal Music Group. And we might expect expert damage models aimed squarely at quantifying market harm.

Practice Tips for the AI-Copyright Trenches

These practical tips can help put you and your clients on the right track, prepared for when the Al landscape shifts, yet again.

1. Start with a Data Audit

   Require provenance disclosures in contracts; price in future indemnification for shadow-library risk. One of the first questions in any Al engagement should be, “Where did the training data come from?” If the answer is, “We’re not sure,” dig deeper. Request a paper (or JSON) trail: licenses, web-scrape logs, shadow-library hits-anything that shows provenance.

2. Memorialize Human Contribution

   Urge clients to keep a prompt log and version history that tracks every post-Al edit. A clean audit trail can influence fair use factor 1 (purpose/character) or rebut claims that outputs are mere substitutes.

3. Take Care with Licensing {and Anti-Scraping) Clauses

   For rights holders: consider “no-Al-training” carve-outs for content licenses and terms of service. For developers, buyers and tool users: demand reps and warranties that datasets were obtained lawfully-and an indemnity if they weren’t.

4. Establish Clear and Compelling Market-Harm Theories

   Meta’s June ruling showed that half-hearted or unquantified arguments for market harm may not make the cut. Make the building of damages models a preliminary preparation step: side-by-side passages, lost-sales projections, dilution calculations, etc.

5. Look Beyond Text

   Song lyrics, voice clones, watermarked images-copyright is not the only hook. Check into trademark dilution, right-of-publicity violations, DMCA section 1202 CMI removal, and unfair-competition claims.

6. Enforce “Prompt Privacy Hygiene”

   Create a policy that prohibits employees from pasting licensed, confidential, or customer data into public models.

The Big Questions

What does “authorship” mean now that models can draft, paint and compose at scale? What does it mean to be a “creator” when the heavy lift of language learning is performed by silicon? How can we serve as effective counselors as courts navigate new claims and theories?

The two June rulings we explored embrace a new and multifaceted identity for text-­ viewing books as both creative works to be protected and as raw statistical fuel, ripe for transformation into something new. This reflects an ongoing shift in global thinking–a shift toward embracing new Al norms. The definition of creator will inevitably evolve, and lawyers who understand both doctrine and data pipelines will be poised to shape the future.

We have voices, and we can use them to insist on sustainable and equitable technology frameworks. We can educate clients and promote the understanding of Al development as it is happening–not just to empowering our clients, but to encouraging mindful tool building and a legal framework that could positively affect creators, developers and the market for generations.