I’m Exempt!

July 23rd 2025

 

Copyright law is territorial. It sounds sensible enough, that each country has its own laws and regulations designed to protect the works of its own authors, artists and creators, right? After all, international treaties standardise copyright protections across jurisdictions and have, in most instances, been fit for purpose. So, what makes AI different? Why would harmonisation of global legislation and regulation be necessary where legal foundations already exist to be repurposed as they have been previously?

 

Well, the purported “fourth industrial revolution” that is AI technology has shaken the framework that copyright was built on. The AI revolution is ubiquitous. The need across jurisdictions to provide legal mapping to progress through such unchartered territory is clear. However, the response of policy makers and the courts is far from harmonious and the pressure to offer hubs for digital innovation has left many in the crossfire of geo-politics, cross-border data flows, forum shopping and regulatory arbitrage. Experts have been unable to confidently offer concrete strategic plans to tackle the cross-jurisdictional battle between innovation and copyright protection. 

 

The solution? Some are advocating for global harmonisation – a more unified approach globally to the legal battle against the AI takeover. However, global unification of legislation and policy addressing AI development and implementation is an unlikely outcome. In any event, countries move slower than tech giants and probably won’t agree on one singular policy, law or approach. Balancing differing opinions and needs regarding innovation and copyright protections globally is no small feat. Especially where scraping high volumes of copyright material to train AI models is concerned. So, what is the current landscape in different jurisdictions? Is it fit to enable innovation whilst offering copyright protection within the bounds of what already exists? Let’s consider two important markets for copyright and tech, the US and the UK, and then a few pieces of insight further afield: 

 

Fair Use

 

Let’s begin in the US with the doctrine of fair use. Codified in Section 107 of the Copyright Act 1976 the US utilises a four-factor test to permit the use of a copyrighted work where the use would otherwise amount to infringement. A specific element of the fair use test that has evolved is transformative use. 

 

First established in the 1994 SC case, Campbell v. Acuff-Rose Music, it has been recently popularised as a potential saving grace for the development and deployment of AI. In 2015 Authors Guild v. Google, the court further clarified that the transformative use exemption applies where the works are used for a purpose different to that of the original work and where the market for the original was not harmed by the use. 

 

Today, AI companies, for example OpenAI, Meta and Anthropic, are relying on these precedents, and others that followed to justify scraping the internet for copyrighted works to use as AI training data. 

 

Let’s first look at the pending Authors Guild v. OpenAI case. OpenAI has taken the position that their use of copyrighted works in the training of their AI models is transformative and did not require the Authors Guild’s permission. They are arguing that they are transforming the works into machine-readable content, specifically, numerical content and algorithms. This, they argue, is not the same as copying. Authors Guild takes the general position of creators, authors and rightsholders alike: training AI on existing copyright work is infringement, licensing agreements should be put in place to authorise use as they are in all other contexts, and unauthorised use without any consent, compensation, and credit undermines creatives, their industry and livelihoods. The New York Times was the first of many additional plaintiffs to follow the Authors Guild, alleging that OpenAI has used their material to train. The NYT case – perhaps the most well-known – showcased that NYT articles used to train OpenAI models could be seen in outputs that mimic the NYT’s writing style, produce blocks of text from NYT articles, and even hallucinate blocks of text falsely attributed to the NYT. OpenAI, in response, famously and pithily argued that: “No one–not even the New York Times–gets to monopolize facts or the rules of language.” 

 

This is true; you can’t police words. And why would we want to scare people into using only a certain vocabulary? But expression, the expression of ideas and the incentivising of creation, that’s what copyright is designed to protect. The question of market substitution returns. The work produced must not in the eyes of copyright be a substitute for the original. There’s no denying that the application of copyright law and indeed fair use is tricky, unpredictable and dependent on the facts of each case. Not what you want to hear when you are looking for clarity!

 

This brings us to the three most recent decisions on summary judgment that reveal differing approaches to employing a ‘fair use’ exemption in the context of AI training: 

 

• In Ross v. Thomson Reuters the ‘fair use’ exemption was flat out rejected. The argument? Well, the case’s facts didn’t satisfy requirements: The copied work was used for the same purpose as that of the original. Particularly, the court noted there was market harm as the secondary work was, in effect, a market substitute for the original. 
• Conversely, in Kadrey v. Meta, transformative use was found to have occurred on summary judgment and a fair use exemption was granted. But interestingly, Judge Chhabria carefully noted that market harm or dilution could be found in other AI cases and it seemed that he granted the exemption somewhat reluctantly. The decision marks the importance of evidencing market harm and it was clear that the ruling was heavily impacted by a failure by the plaintiffs to present evidence of harm. Somewhat surprising given the pirate content that was consumed to train Meta’s LLM, Llama.  
• And then came Bartz v. Anthropic where Judge William Alsup commented that the use of the works in training was “exceedingly transformative”. But, and it is an important but, the fair use exemption only applied to physical books that Anthropic had lawfully purchased, and not the pirate copies it had also ingested. That part is now set for trial on damages as a class action later in 2025. The numbers could be very large indeed and it has prompted Anthropic to seek new lawyers and presumably new arguments.

 

These decisions punctuate the excruciating reality for innovators and creators alike, that the ‘fair use’ exemption “is fact dependent”. Previous cases are unreliable parameters for innovators. To ensure adherence with the law in day-to-day business there is no one-size-fits-all option. The courts will be looking at these issues for years to come.

 

Fair Dealing  

 

Is the UK “fairing” any better than the US with its principle of fair dealing? The fair dealing exemption is limited to a closed list of permitted purposes within the Copyright Designs and Patents Act 1988 Sections 29, 30 and 30A. Generally, we can think of fair dealing as protecting uses such as research, and criticism or review. Unlike the EU, English law does not currently encompass text and data mining. Fair dealing, like US ‘fair use’ relies on the resulting work not being a substitute for the original in the market. Additionally, it also requires that the amount of the work used is reasonable. Pro Sieben v. Carlton 1999, clarified that fair dealing is an objective test, focused on the impact on the audience of the secondary work. Not the intention of its creator. Like in the US, the intention of the author of a secondary work is irrelevant. 

 

Section 29A is probably the most useful part of the 1988 Act, as it relates to making copies of works for text and data analysis. That’s what AI developers say they use works for. Do we have a solution now? Well, the short answer is no. We can’t just apply Section 29A to AI’s use of copyright works to create an exemption, as it is intended to apply to non-commercial research only while, as we all know, the open-source origin stories of AI have been supplanted by subscription models. 

 

Why not expand the scope? That has been a consideration in the UK since 2023, where the UKIPO made a proposal for an all purpose exemption to text and data mining expanding the current section 29A exemption. Attempts to bring this into reality have not worked…..yet. The UK Government, however, has passed the Data (Use and Access) Act 2025 without amending copyright law.  And now, the Government has nine months to carry out an economic assessment and make recommendations. 

 

Getty Images v. Stability AI is the first case in the UK assessing fair use in the context of AI. It has brought up questions regarding input data used for training and the output created by their AI models. However, recently central claims to Getty’s argument (regarding the use of works as input data and the infringing nature of the output generated) have been dropped by Getty. 

 

But why? Remember, copyright is territorial. Well, Getty has struggled to produce concrete evidence to show that the training of their AI system was carried out in the UK. The evidence they have is limited to acts performed outside of the UK. And their output claim? Stability AI has already put in place measures to prevent the use of prompts that lead to the output creation that Getty complained about. 

 

So alas, we get no further in our search for concrete answers about the application of fair dealing and exemptions to AI and we will have to wait for a decision from the High Court. However, Getty has retained claims against Stability for importing Stable Diffusion into the UK, trade mark infringement and passing off, but that’s for another day. 

 

Other Horizons – What else is out there? 

 

The EU has stepped up and offered us the AI Act, alongside the Directive on Copyright in the Digital Single Market 2019 (DSM), and the InfoSoc Directive 2001. But as we have come to expect from Europe what is on offer is a more prescriptive risk-based model. The DSM offers (1) an opt-out scheme for rightholders that don’t want to allow developers to use their “lawfully accessible works for…text and data mining”, and (2) the EU AI Act demands a level of transparency in the data used to train models. So, it’s definitely not the most appealing to developers and risks regulatory arbitrage when it comes to AI training and litigation forum shopping when legal issues do indeed arise. Let’s look at some of the possible reasons for these impacts. 

 

Regulatory arbitrage involves the exploitation by companies of regulatory differences or ‘loopholes’ across or within jurisdictions, specifically, to reduce regulatory burdens or compliance costs. In the context of the EU, there is a clear risk that AI models are likely to be trained in more permissive jurisdictions offering more relaxed regulation. Why? Well, where strict compliance isn’t a factor you need to worry about, there are no hoops to jump through or costs associated with accessing the training data, interest in that jurisdiction as a place to train and develop AI skyrockets. Why wouldn’t companies capitalise on general ease of access to training material and lower development costs? Though legal, regulatory arbitrage is ethically questionable. It leads to unfair distribution of opportunities and unfair competition in markets globally, including, potential financial losses for creators that cannot then benefit from remuneration where their works are exploited outside of their copyright jurisdiction. 

 

Now to forum shopping – the practice of selecting a specific legal jurisdiction or court with the hope of gaining a more favourable ruling. AI companies are likely to choose a jurisdiction with more permissive legislation, regulation and precedents to litigate. With attitudes and legal positions being so polarised globally, forum shopping is the natural result.  Not only does forum shopping reduce opportunities for the development of a more harmonised, unified and ethical global legal landscape. But, the profit over ethics problems can, and likely will, result in deployment of AI systems that are less responsible. As an example, less stringent oversight and regulation could result in amplification of biases found within training data or the algorithms themselves and the biases or unintended outputs could be perpetuated absent intervention.

 

An alternative is an all-purpose exemption, which is in play in Japan and Singapore. Where the distinction between “works as works” and “works as data” has a foothold in legislation. Japan and Singapore’s Copyright Acts at Article 30-4(ii) and Section 243, respectively, offer a complete exemption to text and data mining. This is attractive to developers of AI and innovators. The central token of this fair use ideology stems from a notion that if the original work is not being copied or used for enjoyment purposes then the use does not offend what “copyright fundamentally seeks to protect” (Edwin Tong SC – incumbent Minister of Law in Singapore). Of course this narrative is widely adopted by AI developers globally.

 

There is a global acknowledgement of the use case for text and data mining exemptions. The works are used for data analysis and not enjoyment after all. But what is not clear, and perhaps never will be, is a road map to harmonising and offering a perfectly balanced cookie cutter solution to the needs of innovators and rightsholders.

 

It’s All Fact Dependent

 

So, flexibility comes with uncertainty. As we saw in the cases exploring fair use and fair dealing judges will all too often say “it is fact dependent” to help explain the differences in decision making. And if we achieve certainty through the law we lose flexibility and may find ourselves tied up in more knots. 

 

In the near future it will be up to each jurisdiction to decide, while AI companies may consider uncertainty and resulting future litigation as a cost of doing business. The US has a blueprint in place with transformative use and the fair use doctrine from which it can evolve. It is clear the UK’s copyright law requires an overhaul to aptly address the issue of AI, but any changes will take time. The EU has decided to act now, legislate, and iterate later, as it navigates gaps and inconsistencies with the real needs of creative industries and innovators. The EU usually leads the charge and has what is commonly known as the “Brussels Effect”. But the current political climate has changed the game with a “business first” mantra. It will be interesting to see if other jurisdictions end up permitting innovation without restraint like Japan and Singapore or following a risk-based approach like the EU or  South Korea. Though, even when providing risk-based approaches we can see differentiation between jurisdictions. The EU’s approach being stricter, more rigorous and decentralised when compared with that of South Korea. Which, offers greater flexibility and a more centralised approach, focused more on encouraging risk assessment over prescribing it. Opinions are varied, reactions are varied and the stakes invariably high. 

 

Like so much of AI there is no clear-cut answer and once again the law is playing catch-up with technology. We are watching to see how the jurisdictional and territorial limitations on copyright enforcement will reconcile with the omnipresence of AI and cross-border data flows. This topic will remain top of mind for experts in this space for years to come.