The productivity omission

When it was released last month, the Productivity Commission’s interim report, Harnessing Data and Digital Technology, caused a furore amongst news publishers and the creative industries. The point of most contention was the statement that it was considering whether, in order to take advantage of the potential productivity gains of generative AI, Australia should introduce a copyright exception for text and data mining (TDM).   

The report calls for feedback on the issue, including on the question of what types of uses should be considered fair. But with responses to the report due next week, there is much to be considered about whether we should justify an exception on the basis of productivity gains when there is good reason to doubt whether those gains will be as transformative as AI companies would have us believe. Particularly when this might be at the expense of news, which has clear public-interest value. As AI experts Arvind Narayanan and Sayash Kapoor write, ‘the history of AI is littered with overoptimism about its capabilities and utility’. It may provide incremental gains, but nothing worth betting the future of the publishing and creative industries on. 

What is clearer is that requiring AI companies to remunerate producers for training AI on their content would present a potentially insurmountable barrier to the commercial viability of generative AI. Anthropic’s decision a fortnight ago to settle a class action over the use of pirated books to train its AI models was driven by existential fears. If successful, the action could have resulted in damages topping US$900 billion, well exceeding its annual revenue of $5 billion. Tellingly, Anthropic, like other notable, pure AI companies, makes no profit at all. But we should not expect publishers and creative industries to provide their work for free so that AI companies can profit. 

Indeed, this principle underpins TDM exceptions to copyright that have been established in other jurisdictions, many of which rule out commercial use. But the Productivity Commission underplays this in the survey of TDM exceptions that it puts forward in support of establishing one in Australia. It notes, for example, that there are two TDM exceptions in the EU’s Directive on Copyright in the Digital Single Market, one for scientific research and the second for general use, and observes that the recent case of Kneschke v. LAION ‘endorsed the view that the TDM exception extends to cover AI training’.  

Yet this decision has been criticised. As a recent report for the European Parliament argues, applying TDM exceptions to generative AI stretches the purpose of the directive, which emphasises the need for a ‘well-functioning and fair marketplace for copyright’. Importantly, the general-use exception permits reproductions and extractions of lawfully accessible works. As European copyright scholar Eleonora Rosati argues, this does not mean the same as publicly accessible. Just because the content is on the internet does not mean AI companies are free to treat it as their own.  

The Productivity Commission also argues that, because TDM is a non-expressive use of copyright material, a TDM exception would reflect the principle that ‘copyright law protects the expression of ideas and information and not the information or data itself’. But it is not clear that generative AI is non-expressive in the relevant way. As the European Parliament report argues, generative AI does not merely extract information or identify patterns in its training, but also synthetically reassembles expressive content. Similarly, the US Copyright Office draft report on generative AI training found that the expressive content produced by generative AI is likely to go beyond the boundaries of fair use to the extent it competes in the same market with the copyrighted works it is trained on.  

A further omission from the Productivity Commission report is acknowledgement of the potential for unfettered access to online content—particularly public-interest journalism—to undermine the supply of the quality data that generative AI models need to be useful. This may be, in part, motivating their deals with news companies. But, as we explore in our recent report on GenAI and Journalism, those deals are likely to be on their terms.