The legal landscape for artificial intelligence is entering a period of rapid consolidation. With President Donald Trump’s executive order in December 2025 establishing a national AI framework, the era of conflicting state-level rules may be drawing to a close.
But this doesn’t signal a reduction in AI-related legal risk. It marks the beginning of a different kind of scrutiny—one centered not on regulatory innovation but on the most powerful legal instrument already available to federal courts: copyright law.
The lesson emerging from recent AI litigation, most prominently Bartz v. Anthropic PBC, is that the greatest potential liability to AI developers doesn’t come from what their models generate. It comes from how those models were trained, and from the provenance of the content used in that training.
As the federal government asserts primacy over AI governance, the decisive question will be whether developers can demonstrate that their training corpora were acquired lawfully, licensed appropriately (unless in the public domain), and documented thoroughly.
Inputs Are Key
To date, no US court has held that an AI model’s outputs become infringing derivative works solely because it was trained on copyrighted content. On the other hand, simply using an AI program to copy and incorporate certain copyrighted elements is a different issue and would constitute infringement.
Courts have focused on two settled propositions. First, training on lawfully acquired materials can qualify as fair use. Both Bartz and Kadrey v. Meta Platforms, Inc. held that using legally obtained books in large-scale training is “quintessentially transformative,” a major factor courts consider in determining whether a use is “fair use” or simply infringement.
Importantly, the models don’t reproduce or appropriate protected expression at all—the portion of a work that copyright law actually protects. Instead, in the training phase, they learn statistical relationships among words. Where no protected expression is taken, there is no infringement, and the question of fair use never arises.
Second, fair use collapses when the underlying material was unlawfully obtained. This distinction is now the doctrinal center of gravity. When training data includes pirated books, scraped content of dubious provenance, or copies acquired outside a licensing framework, the fair-use defense evaporates. What remains is straightforward infringement of the exclusive right to reproduce under 17 USC Section 106(1) and the exclusive right to create derivative works under 17 USC Section 106(2).
Once unlawful acquisition is established, the Copyright Act’s punitive feature, statutory damages of up to $150,000 per work for willful infringement, comes into play. When training datasets contain hundreds of thousands of such works protected by copyright, the liability exposure becomes existential.
Bartz as Blueprint
Judge William Alsup’s ruling in Bartz bifurcated the case with precision: Training on legally purchased or licensed works qualified as fair use, while training on pirated copies from shadow libraries didn’t. The court found Anthropic’s use of purchased books “exceedingly transformative,” supporting fair use. However, its downloading and retention of over seven million pirated books weighed against fair use, leading to a damages trial.
The case’s real significance came when Alsup certified a class of 482,460 copyright holders whose works appeared in datasets allegedly downloaded from shadow libraries. This transformed modest damages into an existential threat: potential liabilities exceeding $360 million at the minimum and nearly $72 billion at the maximum. Despite winning the transformative use argument, Anthropic settled for $1.5 billion, the largest copyright settlement in history.
Plaintiffs’ firms now openly describe this approach as the “shadow library strategy.” Track the training data. Identify unlawful copies. Use class aggregation. Invoke statutory damages. Repeat. Nothing in existing law prevents this strategy from being replicated across the industry.
Executive Order’s Impact
Trump’s executive order has been widely described as deregulatory, but this characterization misses its structural consequences. The order recentralizes, rather than diminishes, AI-related legal risk.
First, a unified federal standard means federal courts will now dominate AI enforcement. The order’s creation of a Department of Justice AI litigation task force, its directive to the Department of Commerce to identify conflicting state laws, and its conditioning of federal funding on state compliance ensure that future AI disputes will migrate exclusively to federal forums.
Second, copyright law is already fully developed and used in federal courts. The Copyright Act provides strict liability, statutory damages without proof of harm (provided that certain formalities are completed), class-action mechanisms enabling mass aggregation, and a fair-use doctrine that only potentially applies where there was lawful acquisition of the copyrighted work. As agencies such as the Federal Communications Commission and Federal Trade Commission move toward federal disclosure and provenance requirements, developers won’t merely need to avoid infringement; they will need to prove they did.
Third, federal preemption will amplify intellectual property exposure. With state AI laws increasingly swept aside, the one framework that remains untouched—and increasingly central—is patent and copyright law, which is exclusively in the federal legal arena.
In short, the executive order clears the field for copyright to become the dominant regulator of AI.
Licensed Training Data
Fair use isn’t a blanket excuse for opaque or opportunistic data acquisition. It protects only those who start with lawfully obtained copies. It can’t cure pirated inputs, scraped datasets of unknown provenance, shadow library corpora, or mixed datasets that defy provenance analysis.
Licensing, by contrast, is a complete defense to the reproduction right—the right most implicated in AI training. A model trained on licensed data forecloses the shadow library strategy entirely. It eliminates statutory-damage exposure tied to unlawful acquisition, preserves fair-use arguments for transformative training, satisfies provenance requirements likely to emerge under federal standards, and prevents class aggregation by removing the predicate infringement.
Licensing isn’t an innovation tax. It is the AI industry’s only scalable legal shield.
Outlook
Trump’s executive order may limit the regulatory chaos created of competing state mandates, but it won’t eliminate AI risk. It will shift the arena. With national preemption will come federal scrutiny, and with federal scrutiny will come increased reliance on copyright.
The industry’s stability will turn on whether companies can demonstrate that the foundation of their models rests on lawfully obtained and licensed data.
This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law, Bloomberg Tax, and Bloomberg Government, or its owners.
Author Information
Michael G. McLaughlin is principal in Buchanan Ingersoll & Rooney’s government relations group and co-leader of the firm’s cybersecurity and data privacy practice.
David Gurwin contributed to this article.
Write for Us: Author Guidelines
To contact the editors responsible for this story: