Trilogue Tensions: Shaping the Future of AI Regulation in the EU

The European Union’s (“EU”) proposed regulation on harmonised rules applicable to artificial intelligence (“AI”) systems (“EU AI Act”) has been going through the Trilogues, meetings where the representatives from European lawmaker institutions – the Parliament, Commission and the Council discuss their positions to reach an agreement on a common position, for months now.

With the European elections in June 2024 approaching and different regulatory steps being taken in other parts of the world, including major AI powers such as the US and China, the EU seems to be under pressure to adopt the EU AI Act into law by the end of 2023. However, the closer the finish line becomes, the more visible the controversies seem to get. This, naturally, raises questions regarding the path ahead of the EU AI Act.

In this article, we outline the regulatory procedure the EU AI Act is currently going through and the latest updates from the ongoing Trilogues.

Key Takeaways

• The European Union’s proposal for a regulation on harmonising rules applicable to AI systems is going through Trilogues, a series of meetings during which European lawmakers come together to reach an agreement on a common position.
• The Trilogues have been ongoing since June 2023, and the approaching European Elections that are scheduled to take place in June 2024 are increasing the political pressure to reach an agreement before the end of this year.
• The resurfacing of the divergence between the Council’s and the Parliament’s approach to the regulation of foundation models also raised some concerns as to the future of the proposal, which resulted in the addition of further meetings to the schedule.
• With the pressing urgency to reach a political compromise by the end of this year, these growing divergences may result in some of the most contested elements of the legislation being partly or wholly left out of its final iteration, with their integration into the regulatory framework left to the proposed AI Office, or future harmonised standardisation initiatives provided by the AI Act.

How does the law-making in the EU work?

According to the EU’s ordinary legislation procedure provided under Article 294 of the Treaty on the Functioning of the EU, which is commonly referred to as the “codecision procedure”, all three major European institutions must cooperate. In principle, the European Commission (“Commission”) prepares a proposal, and the European Parliament (“Parliament”) as well as The Council of the EU (“Council”) decide the final outcome. There are three main potential phases of the ordinary legislative procedure, as shown in the figure below:

Figure 1The three phases of ordinary legislative procedure (Source)

As can be seen from above, legislation does not have to go through all three steps and can be adopted as an act even during the first phase.

Where is the EU AI Act in the legislative procedure?

Regarding the EU AI Act, the Commission has published its proposal on 21 April 2021. The Parliament took some time before its first reading. In between, the Council did something that is neither present in the phases above nor uncommon. It published its “general approach” on 25 November 2022 and unanimously adopted it on 6 December 2022, proposing its own amendments to the Commission’s proposal, which is a step taken to expedite the legislative procedure and act as a hint to the Parliament on the Council’s position prior to the Parliament’s first reading.  Finally, on 14 June 2023, the Parliament adopted its position after the first reading.

When the Council’s General Approach and the Parliament’s position are compared, it is possible to identify numerous significant differences between their stances; both diverge significantly from the Commission’s original proposal as well. At this point, rather than going through the formal steps that may result in the rejection, the European lawmakers started to negotiate through an informal set of meetings, aka the Trilogues.

What is a Trilogue?

Depending on the complexity of a given legislation or the existence of conflicting approaches, particularly between the Parliament and the Council, the phases delineated above may reach an impasse, resulting in the rejection of the proposed legislation. In order to reduce the number of such rejections and to provide a forum for discussion, the Parliament and the Council informally come together under the mediation of the Commission, and these meetings are referred to as Trilogues.

A Trilogue may be initiated at any stage of the legislative procedure. Depending on the nature of the meeting, there may be two different types of Trilogues: political and technical. Technical meetings are where detailed issues are discussed, whereas political meetings are where a high-level political agreement is reached based on the outcome of the technical meetings. In the case of an agreement during Trilogues, the ordinary legislative procedure continues from where it was left off, but in a manner that is much quicker and smoother.

With respect to the EU AI Act, Trilogues started immediately after the first reading and the subsequent adoption of the position of the Parliament. In fact, the first Trilogue meeting took place on the very same day as the adoption of the Parliament’s position, 14 June 2023. As of the date of this writing, there have been four political Trilogues with many more technical meetings in between. The next Trilogue, which is the last one this year, will take place on 6 December 2023.

The forthcoming Trilogue holds significance for multiple reasons. The Council's rotating presidency is set to transition from Spain to Belgium; Spain has expressed a keen intention to conclude the EU AI Act within their tenure. Concurrently, policy measures emanating from the United States, the United Kingdom, and the G7 are mounting diplomatic impetus on the EU to finalise an accord on the EU AI Act as soon as possible if the EU is to pioneer AI governance.  Moreover, there will be the final legal drafting of the text as well as the translations after the agreement, meaning that there may not be sufficient time for these before the European Elections in 2024 approaching if an agreement is not reached in this meeting. Last but not least, the recent surfacing of the controversies surrounding some of the most critical issues, primarily the regulation of foundation models, rendered the output of this Trilogue pivotal for the fate of the EU AI Act.

What are the problems that need to be resolved?

There is agreement on some elements of the EU AI Act. By 1 October 2023, according to Kai Zenner, Head of Office and Digital Policy Adviser for MEP Axel Voss, an agreement had been reached regarding the following provisions:

1. Articles 8-12 regulating some of the requirements for high-risk AI systems, with the exception of Annex IV.

2. Articles 16-27 regulating the obligations of providers as well as other actors, with the exception of Articles 23a and 24.

3. Articles 30-39 regulating the notifying authorities and notifying bodies.

4. Articles 40-50 regulating the standards, conformity assessment, certificates, and registration, with the exception of Articles 41 and 47.

5. Articles 53-55 regulating regulatory sandboxes, with the exception of real-world testing.

6. Articles 51, 60 and Annex VIII regulating registration and information to be submitted upon the registration, with the exception of digital gatekeepers.

7. Articles 61-68 regulating market surveillance and post-market monitoring, information sharing, with the exception of access to source code and partially Article 68a.

8. Article 70 regulating confidentiality.

9. Articles 71-72 regulating penalties.

10. Articles 73-74 regulating the delegation of powers.

11. Articles 75-85 on final provisions, with the exception of Article 82b and partly Articles 83-85.

The most controversial issues that needed to be resolved until late October had been the prohibition of remote biometric identification for law enforcement purposes, the balancing of the enforcement between the EU and the Member States, and the regulation of foundation models. There were reports indicating that there was an agreement on the classification of high-risk systems as well as the tiered approach to the regulation of foundation models during the last Trilogue on 24 October. Nevertheless, it was later reported that the negotiations hit a roadblock due to the political objection from Germany and France to the tiered approach to foundation models. Quite recently, it was also reported that France, Germany, and Italy agreed on a “mandatory self-regulatory regime” for foundation models due to the concerns that a stricter regime may hamper the development of European AI firms such as Mistral and Aleph Alpha. In the meantime, the Commission tries to break the deadlock with a compromise text reviving the tiered approach to foundation models.

What to expect?

In light of the above, the Trilogue on 6 December 2023 will certainly be crucial to the ultimate success of the EU AI Act. The political agreement between France, Germany, and Italy does not guarantee approval from the rest of the Member States or from the Council, but it certainly increases the likelihood of political pressure from the Council towards an agreement. The pressure from the EU big three is apparently impacting the technical trilogues, with the latest technical trilogue on 27 November being inconclusive. The next meeting is of the Committee of the Permanent Representatives of the Governments of the Member States to the European Union (COREPER), the main preparatory body of the Council of the EU, and will take place this Friday.

Should a resolution on the regulation of foundational models or other critical issues remain unresolved, a political agreement might still be reached on a text that sidesteps these contentious matters, delegating them to proposed standardisation initiatives (Article 40) and to the AI Office (Article 56) to be established by the EU AI Act in order to preserve the political progress so far and push the draft to be adopted before the European elections. Yet, this raises the question of the efficacy of an AI regulatory framework that leaves its most controversial issues aside. There are strong critics to the approach of the EU Big Three. The effectiveness of such a regulation would hinge on the authority granted to the AI Office and on the normative work to be performed during the grace period. Even excluding foundational models, the EU AI Act is expected to incorporate an extended grace period and a significant dependency on standardisation, which might be regarded more as a pragmatic necessity than a political decision, given the broad drafting of the principles as well as requirements. Indeed, the Commission itself announced that there would be a standard-based early and voluntary compliance scheme under the AI Pact to foster the early implementation of the measures and principles in the EU AI Act before the actual date of entry into force.

Prepare for the EU AI Act

The precise content and the date of entry of the EU AI Act are currently being debated. However, the measures as well as principles embodied within the EU AI Act are universal in nature and similar to those that have been instructed in other jurisdictions’ AI governance rules.

Companies developing and deploying AI will soon have a wave of legal requirements to navigate regardless of where they are located. Compliance is vital to promote safe and ethical AI and avoid large financial penalties and reputational damage.

However, compliance cannot happen overnight. Getting started early is the best way to maximise alignment with emerging and existing laws.

Schedule a call with our experts to find out how Holistic AI can help you with our visionary AI Governance, Risk Management and Compliance Platform, as well as our suite of AI audit solutions.