We have a deal on the AI Act. Finally. To be more precise, it is a political deal.
So, what is exactly the AI Act? It is the flagship law to regulate Artificial Intelligence in the European Union.
The journey started with the AI Strategy in 2020 and the preliminary text from the Commission back in April 2021. Since then, both the Council and the Parliament have been busy shaping their own versions of the proposal. But, as is often the case, trilogues were essential to hammer out a political agreement.
What is a trilogue? Basically, the most mixed-up (and informal) stage of the EU legislative process. But for the nitty-gritty, you might want to check in directly with THE mep-assistant.
This latest trilogue, wrapping up on Friday night, made the headlines. Why? Because of its importance, for sure, and, probably, its marathon length. But, again, we have a deal on the AI Act.
I thought through how to best cover the AI Act in Artifacts. A thorough analysis is probably a bit over my head, while merely reporting the deal would have been extremely boring. Also, there is not a final text yet (!).
Therefore, instead of me talking, today we've got a very, super, special guest. Someone who was in the thick of the negotiations and, most crucially, at the last trilogue of the AI Act.
I am talking about Francesco Vogelezang, Digital Policy advisor for the Greens/EFA group in the European Parliament. He has been extremely kind, and we had an insightful 30-minute conversation.
Needless to say, this Artifacts will not dissect the whole Act. Rather, it will seek to give you a sneak peek into the negotiations, the outcome and what lies ahead for the Act.
Let’s go!
The Interview
Disclaimer: the views reported by the interviewee do not constitute his professional views and do not represent his employer's views.
Lorenzo Ancona: We read in the media about 36 hours of trilogue, split into 2 sessions of 22 hours and 14 hours. Why has it been so lengthy?
Francesco Vogelezang: Actually, I have counted 39 hours. We kicked off on Wednesday at 3 PM with discussions on the regulation of foundation models. There were significant divergences in the positions of the Parliament and the Council, and it took us about 6 hours, until around 11 PM, to find a landing zone. Then, we spent long hours on law enforcement exemption and the bans, as the Parliament had strong reservations about the Council's demands on these issues. After the extended break from 3 PM on Thursday, we reconvened on Friday at 9 AM and worked through the day on the remaining agenda points, eventually reaching a deal by approximately 11 PM.
Yet, it is normal that this process took a lot of time: on some points, there were substantial differences between the Parliament and the Council, and it was indeed challenging to find a compromise under these circumstances.
This newsletter has already discussed the importance of the EU being a frontrunner in AI regulation. A key reason for this is to leverage the so-called Brussels Effect. So, I assume that the AI Act was not just a regulatory challenge but, more significantly, a political one. Would you confirm this assumption? If so, could you comment on the extent of political pressure during the negotiations?
Indeed, the political aspect was substantial, as reflected by the lengthy and complex nature of the negotiations. We were all aware of our shared responsibility in crafting a regulation that could serve as a benchmark for the rest of the world. This is why we stayed there for almost three days: the stakes were just incredibly high.
In the past weeks, Artifacts reported the pushback against the regulation of Foundation Models (FMs) from Germany and France, aimed at protecting their national champions. How did negotiators navigate these divergent interests to reach a consensus on this contentious dossier? Would you describe the outcome as a victory for the EU and a retreat for national interests?
I would say that it is definitely a win. From the start of trilogues, it was never clear whether we would have managed to include foundational models in the regulation's scope. Given quite some resistance from certain Member States, and the degree of attention that the file has received, this outcome was far from guaranteed.
The provisional agreement now regulates foundational models and adopts a future-proof approach by giving the AI Office and the Commission sufficient flexibility to adapt the Regulation’s scope in light of technological evolution. This ensures that the regulation remains technologically neutral and keeps pace with the rapid development of AI.
So, although we in the Parliament had a more ambitious mandate on this specific issue, I think we can be satisfied with the final outcome.
The final text sets a threshold of 1025 for FMs posing systemic risks. Can you explain the rationale behind this specific limit? Additionally, Philip Hacker, a renowned expert in the field, has advocated for lowering the threshold to encompass other models. Do you foresee any adjustments to this parameter in upcoming phases?
The 1025 parameter should be viewed as a reference number. We know that it is very high and it is only met by GPT-4 and a few Chinese models. However, it's meant to serve as a form of symbolic pressure. As I said above, such parameters could be subject to changes as the Commission may update this number via delegated acts in response to future technological developments. Not only that, they can add and remove specific criteria. So, yes, it is entirely plausible that this number will be adjusted, either upwards or downwards, or even complemented, in the future.
Remote Biometric Identification (RBI) was one of the most contentious and controversial topics in the AI Act. As seen in a past release of Artifacts, there was a marked distance between the Parliament and the Council. The outcome is a full ban on RBI with some exceptions, which are yet to be fully clarified. Could you shed light on the discussions surrounding this? From the Parliament's perspective, is there satisfaction with the outcome? Or, as some NGOs claimed in the past few days, these exceptions may not be enough?
As you know, the Parliament's initial stance was a full ban on real-time RBI. This was probably the most tricky point of the negotiations. We were aware that reaching a consensus would require a significant compromise between the two texts given national governments' push for wide exceptions. Considering the initial wide gap, I believe the agreement we reached is quite significant, perhaps even exceeding my initial expectations as we managed to include as many safeguards as possible.
How important is this political agreement, and how important will be the technical work in the coming months? Is there potential for changes in the near future, or can we consider this agreement as finalized?
Echoing the words of Carme Artigas (Spain's Secretary of State for Digitalization and AI, ed) in the press conference, this is not just an agreement in principle, but rather a full political agreement. The technical work will now be crucial to translate what we agreed upon at the political level and make sure that we have the most coherent text possible.
To conclude, what are the next steps?
We don’t have yet an AI ACT per se as it needs to be greenlighted in the Parliament and Council respectively. The political agreement and all the upcoming technical work will be subject to acceptance by the two co-legislators. We will eventually have a final text in Spring, I would say.
Writer’s note: Yesterday, Emmanuel Macron heavily criticized the AI Act, warning that it may hamper innovation and slow down the EU’s competitiveness. He said: “We can decide to regulate much faster and much stronger than our major competitors. But we will regulate things that we will no longer produce or invent. This is never a good idea”. We will, thus, see whether France will accept the present compromise on FMs in the AI Act.
In the Media
What’s in the AI Act, explained.
Google takes on OpenAI with Gemini, its new AI flagship model. Yet, it came under the fire of criticism.
Meta keeps having issues with targeted advertising in the EU.
Beeper broke Apple’s monopoly on iMessage. But Apple wanted to protect its users, so it reinstated the walled garden.
AI is bringing Luigi Einaudi back to life. And Bruno Vespa now speaks Chinese.
The Bookshelf
Unrelated to AI or the EU, I am suggesting “Politics and The English Language” by George Orwell in light of the language shift of this Artifacts (for today). Citing it: “Modern English, especially written English, is full of bad habits which spread by imitation and which can be avoided if one is willing to take the necessary trouble. If one gets rid of these habits one can think more clearly, and to think clearly is a necessary first step toward political regeneration”. A thought-provoking book, at least.
Nerding
When it comes to writing in English, and you’re not a native speaker, QuillBot can be handy. It helps you rephrase sentences, providing several suggestions and tips to make your text sound more bbbritish.