On 13 September 2022, there was on-line discussion entitled “Should the EU Regulate General Purpose AI Systems?”. The webinar was organized by the Center for Data Innovation. The think-tank focuses on issues related to data management and AI policy development in the EU and worldwide.
The discussion panel was attended by:
- Kai Zenner, Head of Office & Digital Policy Advisor to MEP Axel Voss, European Parliament
- Alexandra Belias, International Public Policy Manager, DeepMind
- Anthony Aguirre, Vice-President for Policy and Strategy, Future of Life Institute,
- Andrea Miotti, Head of AI Policy and Governance, Conjecture,
- Irene Solaiman, Policy Director, Hugging Face,
- Hodan Omaar, Senior Policy Analyst, Center for Data Innovation (moderator).
During its presidency of the European Union, France proposed to expand the regulatory area for high-risk artificial intelligence systems, and also introduced regulation aimed at a new concept - general purpose AI systems. These are systems that can perform various tasks of general applicability, e.g. recognize speech or images, generate audio/video tracks, translate, answer questions, detect patterns – and can have a wide variety of final uses, some of which may be unitended. The proposed provisions in the AI Act have sparked a wide public debate, with the legal regulation of these general purpose AI systems having both supporters and opponents.
During the meeting, participants discussed the following issues related to general AI systems:
1. Characteristics of general purpose artificial intelligence systems and why does the EU want to regulate the functioning of these systems?
2. How should the EU legally regulate the use of these systems
3. Recommendations for policymakers in the field of artificial intelligence in the use of general purpose AI systems.
- The discussion began with an explanation of what general purpose artificial intelligence systems are. Remarkably, we can find many different labels used to describe these, such asthe basic system, the basic model or the language model. So far, there has not been a broad social and political discussion about general purpose AI systems and proposals for legal regulation of these systems, e.g. in the AI ACT. It is a system that has many possibilities. It is characterized primarily by a high development potential. It can recognize speech, play chess, pass exams, solve algorithmic tasks, or be part of the code of more complex artificial intelligence systems. The system operation can be compared to the human intellect. By its nature, a given system can perform many activities simultaneously and is classified as a learning system. The algorithms of a given system are in many cases the basis of computer programs and many applications that are widely used. From an economic perspective, these solutions can bring many benefits and spark innovation. However, from a social point of view, attention should be given to the scale of using these systems, their size, their technical characteristics and parameters, as well as the potential for further development on the basis of machine learning. Moreover, the potential for the use of general purpose AI systems faces many difficulties in clarifying and regulating the areas of their use.
- In particular, the essential issues relating to the use of these systems in terms of their purpose, scope, safety, accuracy, degree of compliance, and the division of responsibilities between producers and users should be identified. In addition, appropriate requirements and restrictions on the availability of the use of these systems should be introduced. At the same time, general purpose AI systems should be transparent (open-source code and algorithm), have carefully developed documentation, as well as present a risk assessment and possible consequences of actions already at the stage of programming and testing of these systems. The debate in the European Parliament (EP) on general purpose AI systems is at a very early stage and requires the involvement of a wider range of representatives of social groups to find an appropriate compromise solution. There are many voices and pressures from the high-tech sector to narrow or remove the proposed definitions on AI. The ongoing work on the proposal for a definition of artificial intelligence in the AI ACT in the EP Committees does not fully reflect the applicability of these systems. Therefore, the definition requires changes and clarification, and, above all, distinguishing the specifics of the operation of dedicated (specialized) systems from general ones, as well as the inclusion of an element of supervision of this system by a natural person. During the discussion on the legal solutions introduced in the AI ACT, Kai Zenner emphasized that the proposed provisions on requirements for high-risk AI systems (Article 8, p. 2, 15), issues related to the risk assessment of these systems (Article 9), as well as the human factor taken into account (Article 14) cannot be directly applicable to general purpose AI systems. Therefore, it seems appropriate to add a separate point regulating the general application of AI systems by increasing accountability (new Article 23 (8)) or to define requirements and obligations of concerned systems placed on the market (new Article 55a). The proposed changes require further discussion and the development of an appropriate regulatory solution, taking into account the social and ethical aspects of the use of AI systems.
- General purpose AI systems should, above all, be safe for society and not violate fundamental human rights. The role of the public sector in protecting citizens to whom these systems apply should be strengthened. It is necessary to create appropriate guidelines for the use of such systems and a code of conduct for all service providers, including AI system producers. Transparency and building trust among the public in this regard is an important issue. Regulation on general purpose AI systems should also include sandboxes, taking into account requirements for SMEs and start-ups using general purpose AI systems. Additionally, the created regulations and documents should be developed together with the representatives of different social groups, including social and civic organizations.