President Trump's December 2025 executive order (EO), now being implemented in early 2026, creates a national AI framework that aims to reduce the chaos of state-by-state rules. With ongoing debates over federal preemption, though, states continue to enact their own AI laws. But for areas of clear federal authority, such as copyright law, don't mistake the EO's national framework for an overall reduction in regulatory risk.

According to a recent analysis by Michael McLaughlin in Bloomberg Law, copyright law is poised to become a key federal regulator of AI amid disputes over use of copyrighted works as training data. Indeed, the EO doesn't deregulate AI. Rather, it promotes a unified federal approach that emphasizes federal preemption, positioning copyright law as a key tool to shift disputes involving AI toward federal courts. The EO empowers the DOJ's AI litigation task force to coordinate federal efforts, steering disputes to federal courts. There, legal tools for copyright enforcement like statutory damages and class actions have teeth, while most cases are still driven by private lawsuits from copyright owners.

Key Shifts for AI Companies and Startups

The article analyzes several important shifts for AI companies and startups:

  • Courts are signaling that unlawful use of copyrighted material as training data poses a significant legal risk.
  • Training on legally acquired or licensed data can qualify as fair use, especially for transformative models that learn statistical relationships without reproducing expression, even if use of copyrighted works as training data is not explicitly authorized.
  • Consider Kadrey v. Meta (N.D. Cal., No. 23-cv-03417): Training on legally sourced books was upheld as "quintessentially transformative" fair use, underscoring that both proper provenance and transformative application strengthen defenses against infringement claims.
  • Pirated or otherwise unlawfully sourced data (e.g., from unauthorized copies or shadow libraries) used as training data is where fair use arguments weaken—and statutory damages up to $150,000 per work can scale into existential threats via class actions.
  • The Bartz v. Anthropic case (N.D. Cal., No. 24-cv-05417) showed how "shadow library" claims (involving over 7 million pirated books) led to multi-billion-dollar exposure, culminating in a record $1.5 billion settlement, even for models argued to be transformative.

Emerging Legislative Scrutiny

As an update since this article, emerging legislation like the proposed TRAIN Act would make it easier for copyright owners to subpoena AI training records, adding scrutiny to data practices. This recent news was covered in more detail here.

Bottom Line for AI Builders and Investors

Licensed, well-documented training data is your strongest legal defense.

Read the full piece here.