How The Spread Of Local AI Models Makes Copyright Enforcement Harder

from the the-times-they-are-a-changing dept

The central theme of Walled Culture the book (free digital versions available) is the growing incompatibility between traditional copyright law and the digital, online world. The culmination of that process was the EU Copyright Directive, passed in 2019. The Directive was supposed to be transposed into local laws by 2021, but a year after that deadline, many EU member states had failed to do so. Nor was that a case of mild slippage; a recent report from Communia provides an update on how the implementations are going. Here’s what it found for one of the most contentious areas of the Directive:

the introduction of the press publishers’ right under Article 15 has not been matched by consistent implementation of the mandatory limits intended to contain its scope. Despite the largely prescriptive nature of the Directive, a significant number of Member States have failed to fully implement these safeguards. In addition, some jurisdictions have chosen not to apply existing copyright exceptions to the new right, resulting in a broader scope of protection for press publishers than for other rightholders and further contributing to fragmentation across the internal market.

That fragmentation is deeply ironic, because one of the main justifications for a new copyright Directive was to bring consistency across the EU. As for the even more controversial upload filters, they have proved so difficult to implement that most governments have not even tried to lay down how they should be used:

Most Member States have limited themselves to restating the Directive’s requirement that lawful uploads must not be blocked, leaving the practical balancing of copyright enforcement and freedom of expression largely to platforms and courts. While a small number of jurisdictions have introduced stronger safeguards – such as ex ante protections against overblocking, transparency obligations, and mechanisms to address abusive claims – these remain the exception. As a result, the level of protection for lawful user expression continues to vary across Member States.

Again the much-vaunted consistency that the Directive would bring to EU copyright law is nowhere to be seen. If those failures underline that, as predicted, the EU Copyright Directive has turned out to be a bad law, badly implemented, arguably the arrival of generative AI has made many of its measures completely moot. As Walled Culture has reported, the idea that copyright is largely irrelevant in a world full of AI-generated material – something first suggested on this blog back in October 2022 – is now increasingly mainstream.

But things are still moving fast in the world of generative AI, with yet more profound implications for copyright. A recent post on the IPKat blog explores one of them: the rise of a powerful new generation of AI models that can be run on a personal computer – or even on a smartphone. Many of the latest models coming out of China are not just open source software, but open weight – that is, the models’ numerical values that get set when a model is trained are released, too, so that anyone can download, run, study, and modify them. As a good introduction to this new wave of Chinese AI innovation in Technology Review explains:

If these open-source AI models keep getting better, they will not just offer the cheapest options for people who want access to frontier AI capabilities; they will change where innovation happens and who sets the standards.

One big impact they are likely to have is on the enforcement of copyright, not least in the EU. The IPKat post notes:

The spectrum of engagement with [open source and open weight] models clearly challenges the copyright system. A system, in this Kat’s view, which, until recently, was oriented around individual acts of copying, with platforms cast as new points of interference to bridge technological enforcement gaps. However, [user-generated content] occurring on AI model marketplaces demands a shift in rhetoric and approach. Their ability to redistribute creative agency and control over the tools of cultural production should prompt us to reflect on how copyright law should respond to creativity that occurs through shared infrastructures.

The EU Copyright Directive’s core assumption that the main forums for sharing material would be a few, easily controllable online giants like Google and Facebook, no longer holds. Instead, people are moving to world where millions of people are using the latest generation of open source AI tools collaboratively to generate creations. Those may or may not be based on existing copyright material, but there is no easy way to police that. As the IPKat post points out:

For users running these [new open source and open weight] models locally, they no longer need to pay per request, nor is their data shared with AI companies, and by extension, rightsholders through Article 53(c) of the AI Act.

The world of generative AI is so complex, and moving so quickly, that it is no wonder that even the relatively recent EU AI Act, which entered into force two years ago, is being left behind by the latest developments. And the EU Copyright Directive, which was drawn up nearly a decade ago, is the digital equivalent of the UK’s 1865 Red Flag Act, which governed “self-propelled vehicles”, and required “a man with a red flag was to walk at least 60 yd (55 m) ahead of each vehicle”.

Follow me @glynmoody on Mastodon and on Bluesky. Originally published to Walled Culture.

Filed Under: ai, ai act, eu, eu copyright directive, local ai, local models, open source, open weight