It is impossible to fully and safely regulate complex AI
Gyulai and Ujlaki 21 (Attila, Centre for Social Sciences - Institute for Political Science / University of Public Service. Anna, Centre for Social Sciences – Institute for Political Science / Corvinus University of Budapest, “The political AI: A realist’s account of AI regulation”, Információs Társadalom XXI, no. 2 (2021): 29–42, https://dx.doi.org/10.22503/inftars.XXI.2021.2.3)
Recalling the realist viewpoint of the political sphere, it seems that the only attainable goal is a modus vivendi, which resonates with the idea that an inherent characteristic of the political world is balancing the possibilities of two extremes. History of politics supports this more pessimistic view: occasionally, eruptions of civil war and failed states still embody the brute reality of the Hobbesian state of nature, while the existence of authoritarian and totalitarian dictatures altogether with hybrid regimes are eternal reminders of the impossibility to limit power in a once-and-for-all manner. In light of this reality of the political world, new claims for the regulation of artificial intelligence, more specifically on weak AI, are less promising. Debates on the regulation of AI concentrate on the need to connect principles such as fairness, accountability, safety, sustainability, and social inclusion, among others, to AI governance (for a more exhaustive list, see Hagendorff 2020). Nevertheless, the most discussed issue is transparency, which is among the primary claims for several AI 38 ethics guidelines released by different institutions and companies in the past few years. The current boom in ethical guidelines for AI involves several criticisms concerning the effectiveness of such guidelines based on their potential to implement transparency and other claims effectively. This line of criticism can be divided into three types of argument. The first type challenges the AI guidelines on their extensive list of ethical claims based on their ineffectiveness. This type, which can be called ‘tick-box criticism,’ can be coupled with a proposal of some different approach, for example, virtue ethics (see Hagendorff 2020). The second type, which can be called ‘double standard criticism’, is more sceptical about the possibility of guiding AI and whether full transparency can be achieved at all. This criticism builds on the argument that it would be a double standard to call for higher transparency in AI compared to human decision tools and human reasoning (see Zerilli et al. 2019). The third type of criticism is more focused, what we call ‘specificity criticism’, and argues that current Artificial Intelligence Guidelines (AIGUs) are not specific to AI, but they are simply attempting to gain social control over technology. This criticism also demonstrates that transparency and explainability are claims that specifically concern AI because in such cases there is a possibility of the autonomy of AI. In that case, though, the double standard problem arises (see Héder 2020). These criticisms imply that there is a profoundly political characteristic of AI. On the one hand, there is a relative autonomy inherent in AI that can be understood in a broader sense. It is impossible to regulate in every detail, something that can develop by itself. On the other hand, concerning the expert systems of weak AI, the double standard criticism and specificity criticism correctly acknowledged that it would be an unfair expectation to regulate the decision-making of artificial intelligence in domains where human decision-making cannot be entirely regulated likewise. However, contrary to the double standard criticism, we do not base our argument on the similarity between the obscurity of artificial decision tools and human cognitive processes. Instead, we build our argument on the political characteristic of AI. Using AI as a tool is similar to political authorization: although accountability is the main virtue in politics, it would be unrealistic to expect legislative, executive, or judicial officials to act ‘perfectly’. We can only hope that they behave to the best of their knowledge, and while we usually hold them to account for significant breaches of their power, mostly, we authorize them because authorization is the only legitimate way to create order without slipping into a Hobbesian state of nature or a tyrannical regime.