The blog post below was published on 22.5.2023 in the Constitutional Law Blog, which is the current affairs section of the Finnish Constitutional Law Association. The original text can be found at: https://perustuslakiblogi.wordpress.com/2023/05/22/susanna-lindroos-hovinheimo-tekoalyasetuksen-uusia-vaiheita-parlamentin-nakemykset-laajentavat-saantelya/

The AI Act being prepared in the European Union took a new step on May 11 when the committee work in the European Parliament was completed. Parliament will vote on its position on the regulation proposal in June. After that, trilogue negotiations will begin. It thus appears that the AI Act is coming, and possibly quite soon.

The text has been amended several times since the Commission issued its original proposal. Parliament has added, among other things, a mention of the regulation's objectives to the preamble of the first article:

"The purpose of this Regulation is to promote the adoption of human-centred and trustworthy artificial intelligence and to ensure the protection of health, safety, fundamental rights, democracy, the rule of law and the environment from the harmful effects of artificial intelligence in the Union while also supporting innovation (translation from English SLH)."

The objectives are extremely ambitious. When reading the articles amended by the committees, the question arises whether all this regulation is even possible. The regulation is becoming a broad package that will have significant impacts in many areas of society. Its significance could be compared to the Union's data protection regulation.

The Commission's proposal was criticized from many perspectives, but one problem raised was related to the difficulty of understanding the proposal. This has not disappeared as the legislative process has progressed, quite the opposite. The regulation proposal now shows several tensions. First, definitions – such as the definition of artificial intelligence – are sought to be kept as general and broad as possible. The aim is to bring as many systems as possible within the scope of the regulation now and in the future, regardless of technological developments. On the other hand, many rules concerning artificial intelligence have been formulated in great detail. This has made the whole package not only difficult to comprehend, but also duplicative. On the one hand, regulation is general, on the other hand, it deals with micro-level details. Interpretation will be difficult.

A second tension observed in the regulatory process relates to the fact that regulation has multiple objectives. The aim is to promote innovation and the freedom of internal markets. At the same time, attempts are being made to regulate artificial intelligence in many ways. Thus, promoting innovation and ensuring system security come into conflict. Complex and far-reaching regulation speaks at least to some extent against the development of new systems.

Support for innovation is particularly targeted in Article 53, which would regulate regulatory sandboxes. These are intended to provide a supervised environment that promotes the development, testing, and validation of innovative AI systems for a limited time before their market deployment. The establishment of sandboxes would be the responsibility of member states. Regulation of them has been discussed throughout the regulation's legislative process, and the Parliament's proposal again contains one version of regulatory details. It contains many different rules and requirements and represents a considerably more detailed regulatory philosophy than the Commission's original proposal. One could ask whether such complex regulation is likely to promote innovation.

The emergence of various GPT language models on the market has led to the addition of rules concerning these systems in recent versions of the AI Act. In Parliament, a new proposed article 28b was added to the regulation. It would regulate precisely these systems. Their providers would need to ensure, among other things, that the system is not harmful to health, safety, fundamental rights, the environment, democracy, or the rule of law. In addition, systems must meet several technical requirements. The regulation also requires that systems be designed to be as energy-efficient as possible and take into account, among other things, waste generation. The list of requirements is thus quite long.

The impacts of the AI Act in the domestic context can only be guessed at, but it is certain that new supervisory authorities will need to be established (although authorities will also be established at the Union level). Bureaucracy will increase. Complying with various standardization and other requirements is not easy, so we will also see a demand for new legal and consulting services.

The regulation will apply directly in all member states. In this context, a lot of work is ahead, as all national legislation must be made compatible with the regulation's many rules. For example, the public administration automation package, which just entered into force, may be subject to review. Many other areas already have scattered legislation that falls within the scope of the AI Act.

The regulation leaves very little room for national discretion, but in places there will be some. The Parliament's proposal has added interesting details in places, the significance of which will only become clear when interpreting the final version. For example, in Article 2(5)(c), it would be stipulated that the regulation would not prevent member states from maintaining laws that go further than the regulation in protecting the rights of workers in connection with artificial intelligence. The wording is so far vague that it is difficult to assess the national discretion it would provide.

The regulation will also have economic impacts. It is probably true that automation saves money in, for example, recruitment, healthcare, law enforcement, or education administration. The detailed documentation and reporting required by the regulation does not.

It is certainly commendable that artificial intelligence is regulated. The Union has taken a bold step in undertaking this legislative initiative, which will be followed by much good. However, the AI Act is becoming a broad and complex legislative package, whose interpretation will be extremely difficult. Therefore, it might be necessary to limit the ambition of the regulation so that the regulation also becomes a rational and coherent whole, rather than being all-encompassing.

Susanna Lindroos-Hovinheimo, Professor, University of Helsinki

This article is related to the GenerationAI project funded by the Strategic Research Council, which examines the regulation of artificial intelligence, particularly from the perspective of children's rights.