Unofficial Final Text of EU AI Act Released

On 22 January 2024, an unofficial version of the (presumed) final EU Artificial Intelligence Act (“AI Act”) was released. The AI Act reached political agreement early December 2023 (see our blog post here) and had undergone technical discussions to finalize the text since. It was reported that the document was shared with EU Member State Representatives on 21 January 2024, ahead of a discussion within the Telecom Working Party, a technical body of the EU Council on 24 January 2024, and that formal adoption at the EU Member State ambassador level (i.e. COREPER) will likely follow on 2 February. On Friday 26 January 2024, the Belgian Presidency of the Council officially shared the (analysis of the) final compromise text of the AI Act with Member State representatives – clearly indicating that this text will be put forward for adoption.

The latest rounds of negotiations during technical discussions, according to EU officials, mainly revolved around the threshold for high-impact general purpose AI (“GPAI”) models – which is now determined on the basis of the cumulative amount of computing power used to train these models (expressed in “floating point operation” or “FLOP”) and is now set at 10^25 in the AI Act. In light of technological advancement, the EU Commission has the power to overrule this standard (i.e., consider an AI system to be a “high impact” GPAI despite it not meeting the threshold), and the threshold will be subject to specific review during the legislative review of the AI Act. Further, there currently seems to be a strong consensus that GPAI can be regulated through codes of practice as an interim solution – but that ultimately, the EU shall develop common standards to regulate these models.

Importantly, the EU Commission is already taking steps to implement the AI Act. Under the AI Act, the EU Commission is tasked with setting up an “AI Office” responsible for the regulatory supervision of general-purpose AI models, in particular those with “systemic risks” (i.e. those models which are presumed to have the highest impact on EU society, fundamental rights and values). Although the AI Office will formally be established within the EU Commission (under the administrative structure of DG CONNECT), it is considered independent and will have, among other things, a separate dedicated budget to support its functioning (currently estimated at 60m EUR). The EU Commission issued its decision establishing the AI Office on 24 January 2024.

In recognition of the need for an increase in available computing power and processing capabilities, the EU Commission has also launched on 24 January 2024 its new “AI Innovation Package” which launches a package of measures to support trustworthy AI innovation and market entry by start-ups and SMEs, including a new initiative to provide privileged access to the EU’s supercomputers to innovative EU AI startups and the broader innovation community to train their AI models and foster the development of EU-grown trustworthy AI. This package includes suggested amendments to EU Regulation 2021/1173 establishing the EU High-Performance Computing Joint Undertaking.

In terms of next steps, it is understood that the present unofficial version of the text is the final version for adoption, and that, despite some EU Member States (e.g. France) possibly looking to delay the vote or obtain concessions on the text, there is a high likelihood this text will be adopted in the COREPER meeting on 2 February 2024. According to EU officials, there is “no chance” at this point of the text not getting adopted – whether it be on 2 February 2024 or later that month.

This post is as of the posting date stated above. Sidley Austin LLP assumes no duty to update this post or post about any subsequent developments having a bearing on this post.