UK vs. U.S.: How the Getty v Stability AI Litigation Is Diverging Across Jurisdictions

The UK’s recent decision in Getty Images (US) Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch) has set an important early marker for how courts may approach AI training, copyright, and secondary liability. But the story is far from over. Across the Atlantic, the parallel U.S. proceedings, currently before the Northern District of California, are taking a significantly different path.

Together, the two cases provide a rare glimpse into how the world’s two leading common-law systems may diverge in regulating AI.

1. The UK Approach: Evidence, Territoriality, and Technical Precision

The UK High Court took a characteristically technical and doctrinal approach.

Key Features of the UK Ruling

a. Territoriality was decisive.

The UK court would not consider direct copyright infringement because Getty could not prove that training occurred in the UK. UK law does not capture acts abroad, even by UK companies.

b. No “infringing article.”

The High Court held that Stable Diffusion does not store or reproduce Getty’s works and therefore cannot be an infringing copy under ss.22–23 CDPA. Learning signals and weight changes are not copies.

c. Model architecture matters.

The court engaged deeply with how diffusion models work and concluded that learned parameters are abstractions, not reproductions.

The door to training-stage infringement remains open, but only if evidence of training within the jurisdiction is provided.

The UK’s stance is therefore narrow, doctrinal, and grounded in traditional copyright concepts.

2. The U.S. Approach: Fair Use, Expressive Similarity, and Broad Liability Theories

The U.S. case, by contrast, is focused less on jurisdiction and more on the nature of AI training itself.

Key Features of the U.S. Proceedings

a. Training on copyrighted works is framed as potential mass reproduction.

In the U.S., the act of making temporary copies for training is treated as a core issue, regardless of where servers were located.

b. The debate centres on fair use, not territoriality.

The U.S. courts must determine whether:

• copying during training,
• copying to build feature representations,
• and producing outputs with similar structure or “style”

fall under fair use principles.

The U.S. case engages with expressive similarity and output-level risks.

U.S. plaintiffs argue that even if models do not store images, the outputs may be substantially similar or derivative.

This line of argument has little traction in the UK, where similarity is irrelevant unless an underlying “copy” is stored.

Representative actions are easier. The U.S. class-action system makes it far easier to aggregate large numbers of rights-holders into a single action, unlike the UK’s restrictive CPR 19.8.

Thus, the U.S. litigation is broader, more policy-driven, and more focused on systemic copying, while the UK litigation is more technical, narrower, and evidence-dependent.

3. What This Means for the Future of AI Regulation

The two jurisdictions are now clearly diverging.

In the UK…

• AI developers gain comfort that model weights ≠ infringing copies.
• Rights-holders must prove precise acts of reproduction within the UK.
• The unresolved issue remains:

Is training itself an infringing act?

The UK has not answered, only declined to answer due to lack of evidence.

In the U.S.…

• The Fair Use framework could reshape global AI training norms.
• A ruling against Stability could establish that training = infringement, regardless of where training occurred.
• Output similarity arguments may gain traction.

Globally…

Companies will need to:

• adapt their AI governance to multiple legal theories,
• maintain detailed training-provenance documentation, and
• prepare for licensing regimes covering “training uses.”

The Getty cases illustrate how AI legality will be jurisdiction-specific, not universal.

4. A Simple Way to Understand the Divergence

The UK asks:

“Did you store or import copies in this country?”

The U.S. asks:

“Did you copy the works at all, and if so, is it fair use?”

These are fundamentally different questions.

The UK focuses on mechanics and evidence.
The U.S. focuses on policy and purpose.

The result may be that AI models are lawful in one jurisdiction and infringing in another, based solely on how courts conceptualise “learning.”

5. What to Watch Next

As 2025–2026 unfolds, three developments will shape the global landscape:

a. Northern District of California’s ruling on fair use: A landmark decision with global consequences.

b. Potential UK legislative review: Whether training is formally recognised as copying or exempted.

c. The EU AI Act interplay: Transparency obligations may indirectly reshape copyright norms.

Conclusion:

Two Systems, Two Directions-One Global Challenge

The UK and U.S. cases reveal a widening philosophical gap in copyright and AI:

• The UK prioritises doctrinal precision, territoriality, and technical evidence.
• The U.S. leans toward broader theories of reproduction, fairness, and systemic impact.

For AI developers, this is both a challenge and an opportunity: innovation remains possible, but only with careful governance and jurisdiction-aware legal strategy.

For rights-holders, the message is clear: litigation strategies must tailor themselves to local legal frameworks, what fails in London may succeed in California.

The next 12-18 months will likely define the global rules of AI training, copyright, and the ownership of machine learning itself.