Abstract

Applications of artificially intelligent systems have recently enjoyed increasing popularity, as demonstrated in particular by the presentation of the chatbot “ChatGPT” by the AI research company OpenAI. ChatGPT also demonstrates the potential of AI more generally: Such applications can already be used by private individuals, for example to get support in writing exams or theses, and it seems likely that one day similar AI systems, such as chatbots, could also be used in public administration. This does, nevertheless raise some questions in the public administrative sector: predictive decisions, involving the need for careful risk evaluation, are not new to German administrative law, as is apparent from fields like atomic energy law. Still, it is questionable how far the administration should entrust such procedures of probabilistic risk analysis to AI applications and to what extent this could be legally acceptable?

The analysis in this paper is devoted to answering this question. Part II begins by clarifying some of the key concepts underpinning the notion of artificial intelligence, including ‘digitalisation’, ‘algorithms’, ‘big data analytics’ and ‘machine learning’. The different (supervised and unsupervised) methods of machine learning are considered, as are their modes of operation and selected problems they raise, such as the possibility of bias and lack of transparency and comprehensibility.

Part III then examines in detail the European Commission’s draft regulation on AI (Proposal for a Regulation of the European Parliament and of the Council laying down harmonised Rules on Artificial Intelligence (Artificial Intelligence Act), COM‍(2021) 206 final, 2021/0106 (COD)). This legislative initiative pursues in particular the goal of regulating AI applications and harmonising the regulations on them across the Union. The regulatory scope as well as the area of application and the contents of the draft AI regulation are first shown and possible deficits are discussed. A problematic aspect is the systematic relationship of the draft regulation to other existing EU regulatory acts, such as the General Data Protection Regulation and its Art. 22 GDPR. Following on from the discussion of these problematic aspects, Part IV presents the current draft status of the AI Regulation as well as analysing the (critical) reaction, especially of the Federal Republic of Germany, to the same.

Part V concludes by assessing the implications of the above for use of AI systems in the German public administration, and provides a final evaluation of the draft AI regulation in the latter context. A key issue remains how far it is acceptable for AI-predictions, based on the behavior of a class of persons, to be applied to individual citizens.