A report commissioned by the European Parliament’s Legal Affairs Committee on copyright and AI has concluded that the best way to balance the interests of AI companies and copyright owners is a statutory licensing system. Which is to say, the approach neither AI companies nor copyright owners are rooting for.
A statutory licensing regime would allow AI companies to make use of any existing copyright protected works - such as songs and recordings - when training their models, in return for making payments to creators and rightsholders at rates set by a regulatory authority.
That approach, writes Professor Christian Peukert from the University Of Lausanne in the new report, “ensures wide access to works” for AI companies, which they require, and also effectively “balances the interests of rightsholders, AI developers and end users”.
A statutory licence can be justified, Peukert argues, because without some kind of regulation “creators will demand too much for their data”, while “AI firms will offer too little”. Meanwhile, expecting AI companies to directly license any content that they include in their training datasets results in “high clearance costs” and “too little available training data”.
There has been much debate in recent years, of course, about the copyright obligations of AI companies. When those companies use large quantities of existing creative works - such as music - to form their training datasets, do they need to get permission from the relevant creators and rightsholders?
Peukert’s report outlines the various approaches lawmakers could take. First, provide a wide-ranging copyright exception for AI companies, meaning they don’t need to get permission from anyone to make use of existing works. Second, force AI companies to negotiate licensing deals with every relevant rightsholder, via which they will need to pay licensing fees to access their training data.
The current position in the European Union is somewhere in the middle. There is a text and data mining copyright exception that AI companies reckon they can rely on. However, rightsholders can opt-out of that exception, which means AI companies need to negotiate licensing deals if they plan to make use of any works controlled by opted-out rightsholders.
Neither rightsholders nor AI companies particularly like the EU’s current exception with opt-out system, and statutory licensing is an alternative compromise position.
Under that system, AI companies have easy access to large datasets, but creators and rightsholders also have guaranteed revenue. Though there is then the tricky issue of deciding what AI companies should pay and how that money is then distributed to millions of individual creators and rightsholders across various different strands of the creative and media industries.
Weighing up the pros and cons of the different approaches, Peukert says that a statutory licence provides “broad access to stock of data” and “no need to identify rightsholders”. Plus a “royalty stream” that “restores creation incentives”, thus encouraging people to create and publish new works, which in turn ensures new human created works - and therefore data - to train future AI models.
As for cons, Peukert acknowledges that this approach “requires an independent authority to set royalty rates”, and that the admin of royalty payouts “must be lean”, otherwise there’s the risk that costs associated with administrating the statutory licence use up a lot of the income generated.
The report proposes the use of “automated tools” to process the money, though concedes that an alternative is to use the existing collecting societies that represent different groups of rightsholders.
When the US Copyright Office published a report on the copyright obligations of AI companies earlier this year, it also had a section on statutory licensing, but noted that “most commenters who addressed this issue opposed or raised concerns” about that approach.
“Those representing copyright owners and creators argued that the compulsory licensing of works for use in AI training would be detrimental to their ability to control uses of their works”, it reported.
Meanwhile, “commenters from the technology sector asserted that AI training should not be subject to any licensing regime”, and that even a statutory licensing system would be “logistically unfeasible” and “would result in only meager royalty payments”.
Within the music industry, record labels and music publishers - and groups representing artists and songwriters - are usually critical of any proposals for new statutory or compulsory licensing systems. The fear is that, with statutory licensing, licensees usually end up paying less than with bespoke licensing deals negotiated by individual labels, distributors, publishers and/or collecting societies.
With the major labels now successfully persuading at least some music AI companies to the negotiating table, agreeing presumably lucrative bespoke licensing deals in the process, they'll fiercely oppose any proposals for an AI training statutory licence, in Europe or elsewhere. Indie labels and publishers, and artists and songwriters, will also be nervous about a statutory licensing approach.
That said - given mounting fears within the wider music community that the big AI licensing deals now being announced may end up massively favouring the majors over indies and individual music creators - it might be that a decent portion of that community would be better off with an approach like that proposed by Peukert. Even if the AI sector at large ended up paying less.