The jurisdiction triangle: the US, the UK, and the EU are not fighting the same AI battle

At the FIPP World Media Congress in Madrid last October, I attended a workshop on AI access and attribution where the recurring question wasn’t “how do we license?” It was simpler, and more desperate:

“Shouldn’t we just block any and all bots, and any AI?”

The instinct is understandable. But it’s also a tell: when publishers jump to blocking as the first resort, it indicates they don’t see a way for to them to win, and that there’s only one single, pre-determined outcome.

Of both, there are several.

Last week we looked at how you can use IP to create bargaining power in the age of AI, why this is your best leverage, and what changes when AI can copy instantly.

Part two of this 3-part series frames the fight around three actors (publishers/creators, AI developers, and audiences) and shows how bargaining power changes depending on the legal environment and market structure. In plain English: the same core idea produces different outcomes depending on whether you’re in the US, UK, or EU, and AI firms know it.

The US, UK, and EU form the lion’s share of the global media market. Their differing legal architectures create both litigation risks and strategic arbitrage opportunities for those who understand the mechanics of each region.

Below I provide a map of that triangle, grounded in what the courts and regulators have actually been doing, rather than what anyone wishes they would do.

I often argue that relying on legal outcomes is not a good approach, and that instead we should focus on creating the right incentives which result in the desired outcomes, where everyone wins. This article will show you why I take this position.

In the US, training disputes largely run through Fair Use, which is flexible and fact-specific. That makes outcomes hard to predict, and it shapes behaviour: defendants argue training is sufficiently “transformative”; plaintiffs focus on outputs, market harm, and substitution.

But uncertainty doesn’t mean nothing is happening. Three developments matter.

First: the New York Times’ lawsuit against OpenAI and Microsoft (filed December 27, 2023) puts headline claims on the table: direct and secondary copyright infringement and a DMCA §1202 claim over copyright management information, alongside unfair competition theories. That case matters less because it’s the Times and more because it forces the court to confront how alleged copying shows up in outputs, i.e., whether products can become substitutes for the original work.

Second: in Thomson Reuters v. Ross Intelligence, a federal judge granted summary judgment for Reuters on direct infringement and rejected Ross’s fair use position (as reported by Reuters, February 11, 2025). Whatever happens in the larger wave of media and author suits, that ruling punctures the idea that “training is obviously fair use” as a universal shield.

Third: in the Anthropic v Bartz case, Anthropic settled for a reported $1.5 billion, after it was found they used pirated libraries containing over 400,000 books to train their model, Claude. It emerged in court that they viewed the process of licensing content from each individual rights-holder and author as '“business slog”.

The Fair Use doctrine is currently being stretched by AI labs to justify generative training. However, there is a critical tension point emerging regarding "market dilution"; the theory that AI floods the market with "works of the same kind," depressing demand for the original.

The underlying implication for bargaining becomes explicitly clear: when AI developers believe they can train without permission, your refusal to license doesn’t impose much cost, so training-stage leverage is structurally weak.

Diverging opinion and perspectives amongst the judiciary, however, means that for neither the AI labs nor the creatives, there is no guarantee of outcome.

The UK is arguably the most protective environment for rights-holders.

Rather than a US-style fair use, it is instead closer to a “permission required” baseline for commercial training in many situations, and it flags database rights as a potentially important lever for structured archives.

Under Section 29A of the Copyright, Designs and Patents Act (CDPA) 1988, the UK allows Text and Data Mining (TDM) only for non-commercial research. Crucially, contract terms that restrict this non-commercial use are unenforceable, but for commercial AI training, no default exception exists. This requires AI developers to seek explicit licenses or risk high-stakes litigation.

But there’s a sting in the tail: the UK’s most visible AI copyright fight recently failed to resolve the central training question.

In Getty Images v. Stability AI, the UK High Court handed down a judgment in early November 2025, dismissing Getty’s claim on the basis that the Stable Diffusion models did not contain or store “copies in the model” for that secondary infringement theory.

Two details are crucial:

That matters because publishers often think “UK = stricter = we’re safe.”

You’re not safe.

You may have a different baseline, but you still face the same practical bottlenecks: proving ownership, proving use, and overcoming the enforcement asymmetry between fragmented rights-holders and concentrated AI firms.

If there’s a UK-specific opportunity, it’s not some legal magic wand. It’s that collections and archives (especially verified datasets) may have more legal texture than individual, stand-alone articles, which can strengthen the right-to-exclude when properly documented.

But overall, the UK lesson is that right now AI firms can route around UK courts by training elsewhere, then litigate only over what is imported/distributed in the UK.

The EU is the third corner because it’s neither “wide-open” nor “permission-only.”

The European Parliament has created a two-tier TDM system under the Digital Single Market (DSM) Directive. While Article 3 provides a mandatory exception for research, Article 4 allows commercial TDM unless a rights-holder has opted out via machine-readable signals. The EU AI Act further strengthens this by mandating that General Purpose AI (GPAI) providers publish "sufficiently detailed" summaries of their training data.

In essence, this describes a structured system: commercial text-and-data mining can be allowed unless rights-holders reserve rights/opt out in a recognised way, and newer regulation adds transparency.

The key practical point is simple: in the EU, administrative signals can become commercial variables. This can change the AI firm’s decision criteria from “we can scrape it anyway” to “we may need permission or must route around.”

It’s important not to oversell this though. Opt-outs don’t automatically produce payment, and enforcement still depends on evidence and coordination. But compared to the US, the EU offers more levers that can shift default assumptions, especially for publishers who actually control their archives and rights.

The mistake at the FIPP event wasn’t that publishers asked about blocking. It’s that they presume blocking addresses the real threat.

The crucial point is leverage often attaches to where products are sold and regulated, not merely where training occurred. That’s the logic behind cross-border arbitrage:

This is also why legal “wins” can feel unsatisfying or ineffective. Even when a case goes your way, the industry can pivot: different datasets, different jurisdictions, different product behaviour. So avoid betting the business on a single legal theory or a single court outcome. In reality, incentives shape outcomes.

For publishers, the EU AI Act transparency summary is a golden nugget for discovery. It provides the evidence needed to prove a work was used in training, which can then be weaponised in US or UK courts where such evidence was previously obfuscated.

If you’re a small publisher or a long-form journalist, knowing these regional differences determines whether your strongest leverage is:

Part 3 will be the operational piece: archetype playbooks for independents; what you can do now, with current tools and staff, to build bargaining power without pretending the courts will arrive in time (or even rule in your favour).

Subscribe and tune in next week.