On Wednesday this week (6 December) a final crunch meeting between policy makers in Brussels will be a make or break moment for the EU’s AI Act. EU legislative procedure very rarely generates real moments of high drama, but this is one of them.
The session is the fifth and final official meeting between representatives of the three law-making EU institutions (the European Parliament, Council of the EU and the Commission), in the largely informal trilogue process by which the final text of the AI Act is being negotiated and agreed.
Trilogue on the Act started in June with a stated intention to conclude a political agreement by 31 December. That is when the six-month term of the current Spanish presidency of the Council ends, with a new Belgian presidency taking over in January 2024. The stakes are high because if political agreement is not reached on 6 December then, given the EU’s political calendar for 2024, it is unlikely that agreement can be achieved and a final text agreed, voted on and published before the official recess period begins for the EU Parliamentary elections in June 2024.
In that scenario, formal negotiations on the Act may not be able to continue until September 2024. This would mean discussions restarting in the context of a Hungarian presidency, a slate of new Commissioner appointments and an EU Parliament with a potentially (very) different political composition. The concern is that a changed political climate would mean substantive points in the Act being re-opened, and the whole file then becoming mired in institutional quicksand, with a feasible worst-case scenario being that the Act is abandoned altogether.
In addition, the negotiators are also acutely aware of squandering the EU’s advantage as an early mover in the AI policy space. When the draft of the AI Act was first published in April 2021, it was widely acknowledged as the world’s first comprehensive piece of AI legislation. The Act’s focus on ‘high-risk’ AI systems was seen as a potential template for other jurisdictions to follow. However, with the publication of President Biden’s Executive Order on trustworthy AI on 31 October, the UK-hosted Bletchley Park Summit Declaration on safety risks of Frontier AI on 1 November, and AI laws already in force in China, the EU’s early momentum looks like it could be slipping away. Further delay now could leave the Act looking like an also-ran, ceding ground to these and other legislative approaches with their differing priorities.
What is stymieing the negotiations?
The 6 December trilogue deadline has always felt challenging, with an expectation that a handful of core issues would be taken right up to the wire. However, the discussions moved into a crisis mode after a meeting on 10 November, when Council representatives from France, Germany and Italy unexpectedly hardened their stance on the Parliament’s proposed provisions on foundation models, rejecting them outright. This prompted a walk-out by the Parliament team. Since then, the EU Commission has proposed a different approach (19 November), the Parliament has circulated a working paper (24 November), and the Spanish presidency then presented a revised position to try and find a compromise (28 November). The presidency’s position is reportedly largely based on the Commission’s proposals. It appears to be centred on distinguishing between General Purpose AI models and General Purpose AI systems. It then classifies some General Purpose AI models as systemically risky by reference to quantitative thresholds of compute resources used during pre-training runs, and the amount of EU enterprise users. All GPAI models would be subject to some horizontal regulations, with a more intensive regime applying to those deemed as having systemic risks. Codes of conduct would be published to assist compliance.
Prohibitions and law enforcement
The arguments between the Council and the Parliament over whether remote live biometric identification can be used for some law enforcement purposes, or whether it should just be banned outright, has been the longest running area of disagreement between the institutions. These tensions remain, though reports suggest that the Parliament may now be willing to drop their insistence on a total ban on live use in public spaces, provided that the circumstances in which its use for law enforcement purposes is permitted is very tightly circumscribed, and controlled through layers of safeguards. Some other AI use cases that the Parliament wanted to prohibit, could instead be designated as high-risk. Although both sides have gone on -the-record in the last couple of weeks re-stating the commitments to their respective starting positions, this area seems now to be heading towards a final flurry of horse trading and compromise, provided that Parliament will concede at least some limited live use of biometric identification for law enforcement purposes.
Some other difficult areas remain open, including the powers of the AI Office, the scope of the new fundamental human rights impact assessment and of the national security exemption. Others, notably the set of ‘filters’ added to the high-risk classification process, have reportedly been agreed. Below these more high profile items, is a large amount of minor items which still remain to be definitively pinned down.
Because trilogue is a secretive process, all of the above is only in the public domain largely through the reporting of Euractiv, and updates from Kai Zenner, a policy adviser to Axel Voss MEP who is committed to bringing transparency to the trilogue process. Actual leaked copies of negotiating texts have been hard to come by (which is not always the case with trilogues), so we don’t know the precise detail of what is being discussed.
It seems barely conceivable that, given the pressures to get the AI Act over the line this year, the negotiating teams could fail to reach a political agreement in principle on Wednesday. If they succeed, the timetable would likely be a final agreed text by early February, and a vote in Parliament by early April at the very latest. Publication in the EU’s Official Journal would follow, with the Act applying as law two years later.
Failure on Wednesday would be a disaster for the EU’s digital agenda, and would have a significant impact on the wider international discourse on AI policy making.