The UK government has announced plans to allow artificial intelligence models to be trained on copyrighted content, settling one of the big uncertainties of the current AI revolution ā but the proposal has been criticised by campaigners who worry about the way AI companies already allegedly flout copyright rules.
āThereās nothing balanced about it,ā says , a musician and former executive at AI company Stability AI. āIt will hand most of the UKās creative work to AI companies, for free, letting them build highly scalable competitors to millions of creators.ā He recently organised , signed by more than 37,000 people, calling the use of such training data a āmajor, unjust threatā.
Peter Kyle, the UKās tech secretary, said in a statement: āItās clear that our current AI and copyright framework does not support either our creative industries or our AI sectors to compete on the global stage.ā Kyle claimed the proposed change to copyright law, which will be consulted on, was about ābalancing strong protections for creators while removing barriers to AI innovation; and working together across government and industry sectors to deliver thisā. The UKās Department for Science, Innovation and Technology didnāt respond to a request for comment.
Advertisement
How a regime that aims to serve both ā seemingly competing ā interests will operate is as yet unknown, and will be tricky, says at University College London. Existing UK law allows copyrighted material to be used for research or non-commercial purposes, but as AI has jumped from the lab to the market, it has become unclear how this exemption still applies. āI suspect it means that the text and data-mining exemption is extended to actors with commercial interests, but there is some ability to object to it on a content-by-content basis,ā he says.
Veale suggests that this kind of opt-out regime would be difficult to enforce because of the scale involved ā both on the side of copyright holders who might object to their content being used as training data, and the number of AI companies that would be beholden to follow those requests.
Newton-Rex says opt-out schemes donāt work, because it is impossible to opt out downstream copies of any content. That means, for instance, that if someone illicitly copies a music track or image, the request for it not to be used as training data may not be respected on any subsequent copies. He says by the company he once worked for, Stability AI, suggests uptake of opt-out schemes hovers around 10 per cent or less because of the burden required.
An opt-out approach would echo the , which includes obligations for AI model providers to describe the sources of data their systems are based on and to allow people to remove their data. This is more complex than it sounds, as techniques for āmachine unlearningā that can remove training data from an AI model without the need for lengthy and costly retraining are still in their infancy.
Despite these issues, some believe the UK proposals are a good first step forward. āThe AI industry have, until now, treated the owners of copyright with contempt,ā says at University College London. āSome creators wonāt be satisfied with ālegal certaintyā or a ābalancedā approach when they feel that these AI companies have stolen their property. But if government is able to get AI companies to be transparent about where their data comes from, it will be a substantial victory, and the start of a much-needed process of accountability.ā
at the University of Staffordshire, UK, is similarly sanguine. āI think this is a good start to reining in the complete free-for-all strip mining of copyrighted and other content taken without the makerās consent that the big tech companies have been getting away with in the rush to produce results,ā she says.
But ultimately, any settlement that doesnāt involve licensing copyrighted material, as some AI firms are doing with news providers and other organisations, involves siding against creators. āAs well as making a significant contribution to our economy, the UKās cultural and creative industries are valued worldwide for their influence and inspiration,ā says at Kingās College London. āAny proposal that waters down the UKās copyright laws puts the creative industries at risk.ā
