Guido Westkamp (Queen Mary University of London; University of Münster) has posted The Freedom to Mine and Train Artificial Intelligence Models and the Limits of Copyright and Personality Rights Law on SSRN. Here is the abstract:
It seems, as things stand under the EU AI Act, almost impossible to develop counter-rights based on personality interests (which include copyright) for the sake of human self-autonomy. The current debate on the legitimacy of AI centres on what happens in AI training, and that focus obfuscates the relevant interests since it must rely on assessing, centrally, the duration of copies which evidently says little about the relevant interests (substitutability in economic terms, affected personality interests) and how these can be aligned, if at all. The contribution addresses several problematic issues of applying “black letter” law to AI uses in the context of both copyright and personality rights by first framing the conflicts of interests as one between collective rights to communicative freedoms to be balanced against the rights of authors, or creators more generally. The relevance of training AI models, and here in particular the consequences of the “black box” phenomenon, will be analysed and exemplified by two conflicting decisions from Germany and the UK, concluding that the German decision in GEMA v OpenAI, which concludes that AI training is not permissible – and which very much exposes the deficiencies of a “strictly legal” point of view – is doctrinally misguided in both rejecting the applicability of the data mining exception under Article 4 DSMD and, more importantly, eschewing a detailed assessment of the complex interests at stake by focussing almost exclusively on allegedly unlawful memorisation and perceptibility. It will be demonstrated that an assessment “in reverse”, starting with the question whether the provision of AI models can constitute an “illicit” service under the right of communication to the public before turning to the legitimacy of temporary copies, can produce the diametrically opposed conclusion with ease, which can be brought about by relying on principles orbiting around self-responsibility and implied consent – in short, whether AI can or should be categorised as a necessary and socially acceptable technology which users envisage to remain available. The debate as regards copyright is decidedly pertinent to the status of (highly fragmented) personality rights which lack, across jurisdictions, a clear fundamental rationale when considering cases of appropriation of human image, voice or data. Some emphasis is placed on the apparent inadequacies of the existing legal framework: whether AI training involves the use of personal data at all, whether data privacy – as a fundamental right – effectively offers a right to object, and in how far the current political emphasis on a European data strategy might further restrain the relevance of human claims to self-determination in light of being exposed to generative AI.
Recommended!
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum