• Europe’s AI Act has reached another milestone
• EU’s council of ministers reach an agreement on the wording of the AI Act
• It’s not rubber stamped yet, though 
• Bureaucratic and political niggles are likely in the next stages 
• Negotiations have been described as akin to “herding cats in a soot storm”

The member states of the European Union (EU) have taken another step towards passing into European law the AI Act – unique and groundbreaking artificial intelligence legislation designed to prevent any AI systems or models deemed to pose an “unacceptable risk” from being deployed anywhere in the region and imposing different levels of legally binding obligations and requirements on any deployments categorised as either “high-risk” or “limited risk”.

But while a landmark agreement has been achieved, this time by the Council of Ministers’ Permanent Representatives Committee, there’s still a long way to go.

The EU AI Act was first proposed by the European Commission (EC) back in April 2021 and its provisions and proposals have been scrutinised, debated, re-scrutinised, argued over and kicked around like political football ever since by the Council of Ministers and representatives of individual member states in the European Parliament.

The matter is urgent because AI technology is evolving and morphing quickly and its application in daily life is becoming commonplace, if not yet ubiquitous. Catching up, legislatively, in such a febrile environment will be very difficult and speed is of the essence. However, both the Council of Ministers and the European Parliament must first agree on every word in the text of the proposed legislation and then formally “adopt” it before it can become EU law. It’s a process that has been described as akin to herding cats in a force 10 sootstorm.

The proposed new law is beset, as is usual, by the differences in opinion and approach to the problem of how to control and regulate AI by the different member states as well as the convoluted language (in English at least) used to report what is going on. However, it can be gleaned from various official announcements that a major potential stumbling block has been surmounted as, according to the chairman of the rotating presidency of the EU, currently held by Belgium, “Members of the competent Council of Ministers' Permanent Representatives Committee (Coreper) voted to approve the AI Act at a meeting on Friday morning [February 2].” In a “a milestone, marking the first rules for AI in the EU, aiming to make it safe and in respect of EU fundamental rights” the Coreper has “confirmed the final compromise text found on the proposal on harmonised rules on artificial intelligence”. 

The compromise came after Austria, France, Germany and Italy, pushing their individual member state self-interests, were persuaded to desist from making eleventh hour changes to the report (on the issue of whether or not the AI Act would reduce technical innovation in the EU) before it went to a Coreper vote. France had been adamant on pushing for a less stringent regulation of the big, powerful and established AI models such as Open AI’s GPT-4, that support General Purpose AI systems such as ChatGPT and Bard. Furthermore, the EU’s three biggest economies, Germany, France and Italy, were demanding limitations on regulation that might make life more difficult for European startup AI companies that might, one day, grow to become viable competitors to US AI enterprises.

Moving not so quickly on, the Members of the European Parliament (MEPs) who sit on the two committees within the Parliament that are jointly leading scrutiny of the EU AI Act proposals on behalf of the Parliament itself (one being the Committees on the Internal Market and Consumer Protection and the other the Committee on Civil Liberties, Justice and Home Affairs) have to confirm that they are in accord with the framing and language of the legislation. After that happens the proposals will then be put to a vote of all MEPs in a plenary session in the Parliament, who must agree to their adoption. Thereafter, the Council itself will formally vote whether or not to adopt the text. Simple! 

Whenever, if ever, the legislation makes it onto the EU statute books, a new risk-based regulatory system will then apply across the entire EU. Thus, some uses of AI will be completely prohibited. The prohibitions will pertain to mandatory requirements in the areas of risk management, data quality, transparency, human oversight and accuracy. Meanwhile, “providers and deployers” will face legally-binding restrictions on registration, quality management, monitoring, record-keeping, and incident reporting.

Additionally, regulation will also pertain to general purpose AI models, including the obligation to provide detailed summaries about the content used to train AI models and in the case of “general purpose AI models” deemed to pose “systemic risk” will have to oversee model evaluation and adversarial testing.

– Martyn Warwick, Editor in Chief, TelecomTV