Progress on the AI Act and AI Treaty, but the regulatory work is unfinished.

This article was published by Gregor Strojin, CAI Vice-Chair, on his LinkedIn page. He writes:

The length record-breaking final meeting of the trialogue managed to achieve a political consensus among EU co-legislators on the provisions of the Artificial Intelligence Regulation (AIA) late Friday in Brussels. This is an important step towards the much-needed binding legal framework to ensure the development and use of AI by incorporating risk prevention and the assurance of safety, quality and fundamental rights in technological progress while simultaneously stimulating technological innovation and investment. But the work is not yet finished.

The AIA is a wide-ranging legal instrument that addresses a vast and complex area in which there are many conflicting views: between developers and users, between different branches of business and civil society, but also between different governments, within governments, between branches of government, and of course political groups. Given this, it is understandable that any final result represents many compromises that can fully satisfy only a few.

That said, we currently do not have binding legislation, and many events of the past years, weeks and even days indicate a frontier mentality. Both the Spanish Presidency of the Council of the EU and the European Commission, in its role of “honest broker”, have played a remarkable part in bringing together the different positions of political and other stakeholders. In doing so, the representatives of the European Parliament have notably played their traditional role of protectors of citizens and their rights, insisting on positions to ensure the protection of human rights, including limits to the use of AI in law enforcement (policing) and the transparency of foundation models or general-purpose AI systems.

The political consensus in the trialogue still needs to provide the final text and, thus, all the details and information. The “technical details” that were skipped over during the trialogue meeting are yet to be put into text, and they will be crucial to understanding and analysing all the rights, obligations, and procedures to come. After a series of interim working group meetings through January, the final text of the AIA will have to be endorsed by both the European Council and the European Parliament, which is supposed to have the final say in March 2024. Nothing is agreed until everything is agreed. The timetable is particularly tight as the mandate of the current European Parliament, as well as of the European Commission, is about to expire, and the EU elections will be held in June 2024.

Since June 2023, the European Commission has promoted the idea of a so-called AI Pact, a voluntary mechanism intended to ensure compliance with the envisaged AIA principles during the transition period until AIA enters into force. If, for whatever reason, the final text of the AIA is not adopted by the European Council or the European Parliament by the end of the current mandate of the EU institutions, there is a risk that voluntary mechanisms will become the norm and that the momentum for the preparation of binding instruments will diminish. Given the rapid development of technology, it is questionable whether it will be possible to revive it at the same level in time.

The absence of binding regulation is a form of regulation itself. Self-regulation is already possible but has shown to be ineffective and unenforceable in light of the enormous stakes and interests at play. These will likely continue to direct much of their energy toward derailing or diluting the AIA, as was seen with intense national lobbying on the pretext of foundation models, which almost scuttled the adoption process only within the last month. In an almost serendipitous turn of events, the Open AI drama took place, showing in many ways not only that self-regulation might be insufficient but also that foundation models of such size and user base could be considered at a level similar to that of critical infrastructure. If such efforts succeed, the European politicians, particularly the Commission, will not be able to count an outcome with voluntary instruments alone as a success, despite or even particularly due to the forthcoming elections. Binding regulation is needed to provide much-needed legal and technical certainty, even if that cannot be achieved completely and perfectly.

In parallel with the trialogue, the 8th session of the Council of Europe’s Committee on Artificial Intelligence (CAI) took place in Strasbourg from 5 to 8 December, where the text of a future international treaty for the design, development and use of AI in line with the standards of human rights, democracy and the rule of law is being negotiated.

Based on many principles common with the EU Regulation, the future convention will establish a legal framework to which, in addition to the 46 member states of the Council of Europe, other countries worldwide can accede. In addition to the EU and other European countries, the US, Canada, Japan, Israel, Mexico, Peru and Argentina actively participate in the process, which strives to ensure a globally compatible framework in the field. AIA is a maximum harmonisation regulation, so we can consider it a future implementation of the prospective treaty for the EU27 member states. The work of the CAI is expected to be completed at its 10th Session in March 2024, with the treaty ready for signature in May 2024.

In line with the timetable, the AIA is also expected to be published in the Official Journal of the EU in May 2024. In the meantime, before spring, winter is coming. Nevertheless, in AI regulation development and various technological deployments that we grew accustomed to seeing at an increasingly fast rate, things will remain hot, and the winter will likely be only literal, not metaphorical.

Europe is not alone in preparing or enacting legislation governing AI. China and the US passed relevant legal instruments in 2023, and AI regulation is globally taking shape. The white smoke from Brussels shortly before midnight last Friday is a critical signal to all AI systems’ developers, deployers, users, or those considering becoming one. Stock-taking of their current plans and procedures, mapping them to future rules, employing quality management mechanisms envisaged by the AIA and paying attention to the accompanying standard-setting activities will be necessary to achieve and maintain compliance and competitiveness once they become mandatory. Significantly, there was little political and technical disagreement about this.

Please click on this link to read the original article.

Image credit: Image by masadepan on Freepik

Your account