What Can AI Teach Us? Legal Grounds and Limitations of Machine Learning Regulation

The global AI race in recent years has forced states to reconsider their regulatory approaches to make them more flexible and provide greater support for innovation. Advanced AI systems are becoming increasingly productive and impressive in terms of their capabilities, yet the risks to society are increasing. One of the key problems in the early stage of developing, introducing, and applying AI systems, particularly large generative models, is the availability of data and its selection for machine learning purposes. The world has repeatedly had to deal with situations when the wrong approach to machine learning on the part of developers led to controversial or illegal results that affect the rights and interests of various entities. However, the problems associated with AI are not limited to the availability of data. Other important issues include regulating high-risk AI systems, imposing commensurate requirements on their developers and operators, the appropriateness of labelling AI content, as well as assessments of AI systems to see if they possess any dangerous capabilities. What is the optimal way to determine legal requirements for high-risk AI systems? When should external and internal assessments of AI systems be conducted and who should be responsible for this? Would the mandatory labelling of AI content build greater trust or simply add more bureaucracy?