Powerful AI models that people can download and run on their own computers are making a familiar copyright enforcement plan look badly dated: push responsibility onto large online platforms, then regulate those platforms. Walled Culture argues that the model breaks down when creation happens through open-source and open-weight AI tools on personal devices, beyond the usual checkpoints.
The problem is sharper in the European Union because the bloc’s 2019 Copyright Directive was built around central points of control. The directive was meant to be written into national law by 2021. Walled Culture has reported that many EU member states were still behind a year after that deadline.
A recent report by Communia, cited by Walled Culture, says the implementation record remains uneven. Communia found that Article 15, the directive’s press publishers’ right, has not been matched across member states by consistent limits on its reach. According to Communia, some countries did not fully implement the required safeguards, while others declined to apply existing copyright exceptions to the new right. The result, Communia said, is more protection for press publishers in some jurisdictions than other rightsholders receive, and more fragmentation inside the EU market.
Upload filters left to platforms and courts
Communia also found that member states have mostly avoided spelling out how upload filters should work in practice. The directive requires that lawful uploads should not be blocked, but Communia said many governments have largely repeated that rule without giving platforms much operational detail.
Some countries added stronger protections, according to Communia, including measures meant to prevent overblocking before it happens, transparency duties and systems for dealing with abusive copyright claims. Communia said those examples remain exceptions, leaving users with different levels of protection for lawful expression depending on the country.
That is an awkward result for a directive sold as a harmonizing measure. Walled Culture’s critique is blunter: the law was difficult to implement even before generative AI changed where copying, remixing and production happen.
Open weights move the fight off the platform
The newer problem is local AI. IPKat has examined the rise of open-source and open-weight models that can run on personal computers and, in some cases, smartphones. Open-source software makes the code available. Open-weight releases go further by publishing the trained numerical parameters of the model, allowing users to download, run, study and alter the system.
Technology Review, in coverage cited by Walled Culture, described a wave of Chinese open-source AI development and said that if these models keep improving, they could offer low-cost access to frontier capabilities while shifting where AI innovation happens and who sets technical norms.
For copyright enforcement, the mechanism matters. A platform-hosted AI service can log prompts, meter access, set moderation rules and produce records that regulators or rightsholders may seek. A locally run model does not need a per-request payment system, and user data is not sent back to an AI company, IPKat noted. IPKat also pointed to Article 53(c) of the EU AI Act, saying that local use can keep data from flowing to AI companies and, by extension, rightsholders through that route.
IPKat argued that open-source and open-weight models strain a copyright system focused on individual acts of copying and on platforms as enforcement choke points. User-generated material produced through AI model marketplaces and shared infrastructure, IPKat said, requires a different legal approach.
Walled Culture’s conclusion is that the EU Copyright Directive assumed a world dominated by a few manageable online giants such as Google and Facebook. Local AI models point to a messier one, where many users can generate works collaboratively with tools that may or may not draw on copyrighted material, and where there is no obvious upload gate to police.
This story draws on original reporting from Techdirt.