EU votes for AI copyright disclosure in ‘first of a kind’ act
Transparency rules over copyrighted content used in training data increase risks for foundational AI developers| Landmark AI Act thought to be the first of its kind in the world | Insights from Taylor Wessing, Gowling WLG, Appleyard Lees, Bird & Bird, Marks & Clerk and Potter Clarkson.
EU lawmakers have backed the introduction of more stringent artificial intelligence (AI) rules, which could force companies behind tools such as Chat GPT to reveal the materials used to train their generative AI systems.
The vote which took place yesterday, June 14, could also see a ban on the use of the technology in biometric surveillance, emotion recognition, and predictive policing of AI systems.
The landmark AI Act, which has been in development since 2021, is thought to be the first of its kind in the world.
Following feedback on its initial draft, the European Commission (EC) put forward amendments to its proposed reforms that aim to set a global standard for AI technology.
Yesterday’s vote means that the draft legislation will now inch towards its final stage of being put before member states for approval before it can become law.
‘No pause button’ for AI
Commenting on the development, Thierry Breton, Commissioner for Internal Market of the European Union, said. “AI raises a lot of questions— socially, ethically, economically. But now is not the time to hit any 'pause button'. On the contrary, it is about acting fast and taking responsibility,”
Microsoft, which has called for AI rules, welcomed the development
“We believe that AI requires legislative guardrails, alignment efforts at an international level, and meaningful voluntary actions by companies that develop and deploy AI," a Microsoft spokesperson said.
Meanwhile, OpenAI, founder of generative AI tool ChatGPT, has criticised the regulations, suggesting that it may have to halt operations in the EU if they come into force.
The late mention of copyright to the act is perhaps a reflection of how quickly the technology is developing and a nod to recent litigation against AI tools.
As Fritz-Ulli Pieper, partner at Taylor Wessing explained, different AI Act proposals by the EU Commission, the Council of the EU and the Parliament have evolved and changed significantly over time.
Copyright issues were not an initial focus of the EU Commission, nor of the Council of the EU, in their respective AI Act drafts, he noted.
However, Pieper added, this changed when it was the EU Parliament’s “turn to come up with its own proposal”, which did mention copyright issues.
“This was likely due to the rapid development of generative AI in parallel to the parliaments’ committees’ work on the AI Act. In light of this, it will be highly interesting to see how the EU institutions take into account IP and copyright."
‘Huge practical difficulties’ of AI disclosure rule
A requirement for companies to disclose details of the copyright works used to train generative AI is undoubtedly the biggest IP aspect of the act, according to Matt Hervey, partner at Gowling WLG.
To date, companies have not been obliged to reveal such details when it comes to developing training sets for their generative AI systems—to the chagrin of copyright owners.
Alluding to lawsuits filed against Stability, AI, MidJourney and others, he said: “There are already copyright disputes over generative AI, and this obligation would make it easier for rights owners to identify potential infringements.”
Toby Bond, partner at Bird & Bird, agreed that the main impact on IP is the transparency obligation imposed on providers of foundational models, which compels them to “document and make publicly available a summary of the use of training data protected under copyright law”.
Having to disclose this information, he suggested, could increase the number of third-party claims against developers relating to their use of training data. It could also create “huge practical difficulties if they are required to identify the rights owners for each copyright work”.
“Developers may also be concerned that granular information about their training data could disclose trade secrets around their training process,” said Bond.
“The level of detail required by this obligation will be key and will need to be carefully considered during the trilogue negotiations before the act is finalised.”
Indeed, some expressed surprise that such a radical amendment survived the lawmakers’ scrutiny without any further tweaks.
As Mark Nichols, senior associate at Potter Clarkson, observed: “The most striking aspect of the proceedings is that the AI Act passed without amendment.
“The obligations on foundation models such as ChatGPT to publish the details of their training datasets, and to mark materials generated by such models as ‘AI generated’, has remained.
“There was a debate around whether to classify real-time use of AI to identify individuals via surveillance systems as high risk and therefore forbidden, but this aside, the act passed unamended.”
Unanswered infringement question
Notably, the act doesn’t specifically tackle the question of whether copyright is violated when generative models mine data on the internet for training purposes.
Parminder Lally, partner at Appleyard Lees pointed out that the controversial issue is still ripe for legislators and law enforcers to deal with separately.
Nonetheless, because wording of the act is broad, it will have far-reaching implications for a host of providers and users of software systems which generate content or make predictions and recommendations.
“It will impact AI/software systems in a variety of fields, from healthcare and medical diagnoses to journalism and content creation, to systems that are used to hire, fire or promote people,” she explained.
Striking a more cautious note, Simon Portman, of counsel at Marks & Clerk, observed that the tougher regulations could be compared unfavourably to more lax regimes, such as Japan and the US.
“Algorithms and patent applications will now have to be written with compliance in mind. There is inevitably concern that the practical and financial burden of attaining compliance will make the EU a non competitive environment for AI, especially if the US and Asian approach is more relaxed,” said Portman.
Such challenges, he suggests, could mean that some companies come to find successful innovation prohibitively difficult or, worse, ignore the regulations completely and hope they don’t get found out.
“For example, the speed and complexity with which AI works have made the General Data Protection Regulation cornerstone of transparency in data processing unworkable in many cases.
"It’s difficult to see how the AI Act’s requirement of full disclosure of copyright material relied on by generative AI will be any more practical,” he said.
It’s not over…
Given that companies such as OpenAI have expressed misgivings, Nichols predicts that the act‘s advocates—despite having propelled the draft legislation to its final stages—are not quite home and dry yet.
“OpenAI, the entity behind ChatGPT, has hinted that it may have to consider ceasing to operate in the EU if [the act] is passed and it cannot comply,” he explained.
Whether it will follow through remains to be seen, “but it may well be the case that this debate is not yet over, and that there will be significant lobbying still to come,” added Nichols.
The vote comes as lawmakers around the globe try to grapple with the legal implications and questions surrounding the explosion of generative AI.
This month Australia announced that it plans to unveil more regulations to govern AI, while the UK and US signed the ‘Atlantic Declaration’, a series of mini deals looking at AI, alongside critical minerals, clean energy and security.
Elsewhere, Japan’s government’s newly formed council on AI strategy has raised concerns about the country’s lack of regulation for AI, including the potential risks it poses to copyright infringement.
