Reflections on Copyright and Generative AI

Yesterday (20 March), I spoke at UCL’s IBIL event 'Copyright and Generative AI: Putting input issues under the spotlight'.

Such events remind you that every day is a school day. There are two particular issues I am mulling over:

1. The UK government has to make its next move following the acknowledgment that it will not be possible for AI developers and creators to agree a code of practice to address the relationship between intellectual property rights and generative AI. It is plausible that the UK government will advance a copyright exception for text and data mining (TDM) along the lines of Article 4 of the EU Digital Single Market Directive (under which TDM is permitted for any purpose, but subject to the right holder not having reserved the right to such use). Many of us were expecting that back in June 2022 when, instead, the government proposed much broader TDM exception with no right holder reservation - perhaps the temptation to do something demonstrably different to the EU was too much to resist. I imagine that many big tech entities will have little interest in such a proposal and at least while the substantial volume of cases in the US – in excess of 20 at my last count - have yet to be decided. In many of those cases there is an open question as to whether making copies of copyright works for training generative AI is a ‘fair use’.
    
2. The second issue concerns whether the tokenising of a lawfully acquired copyright work in the course of training an AI amounts to a reproduction. If it does, there could be a copyright infringement. Can the tokens be used to ‘reverse engineer’ a copy? It reminds me of an unreported Oslo District Court decision in 2009 which concluded that an indexing process (analysing and transferring data into an search index) so that words of literary works were translated into numerical codes from which the works could be ‘reassembled’, was a reproduction under Norwegian copyright law. Indexing is not the same thing as tokenising, but if the works can be ‘reassembled’, then ostensibly a further copy is made for which authorisation would typically be required.

We’ll be watching these developments closely and updating in due course. In the meantime, you can subscribe to our AI-related insights here.