MEPs adopted, on Wednesday 14 June, the report by Dragoş Tudorache (Renew Europe, Romanian) and Brando Benifei (S&D, Italian) on the Artificial Intelligence Act (AI Act) (499 votes in favour, 28 against and 93 abstentions) (see EUROPE 13200/1). The European Parliament’s negotiating mandate was also approved by MEPs and the first interinstitutional negotiations (trilogues) began the evening after Parliament’s vote.
“I think we have made history today. We have set the way for the dialogue we must have, which we have already begun, with the rest of the world. From now on, we will be able to build an AI that is responsible in the face of the systemic risks it represents”, declared Mr Benifei following the vote.
EPP amendments on remote biometric identification rejected
Firstly, the text confirms the ban on remote biometric identification in real time and a posteriori in areas accessible to the public. The EPP, which had initially pushed for authorisation of this method, scaled back its ambitions after seeing its amendments rejected in the European Parliament committee on 11 May (see EUROPE 13197/5). At the time, Manfred Weber’s group tabled a number of amendments aimed at reverting to the provisions of the Commission’s proposal, which provided for real-time biometric identification to be authorised in exceptional circumstances, such as the search for a missing person, locating a suspect in a serious crime or preventing a terrorist threat. All amendments along these lines tabled by the EPP group were rejected in the hemicycle (see EUROPE 13180/1).
“Biometric identification was discussed. We have reached an agreement that takes into account respect for privacy for citizens and society, and goes even further. The Commission wanted an exemption for law enforcement agencies. Our report states that this can be used with the explicit authorisation of the courts, so for us it’s balanced”, said Mr Tudorache.
Discussions on the subject had distanced the EPP from the Renew Europe, S&D, Greens/EFA and The Left groups. However, the co-rapporteurs were confident that these divisions would not become apparent during the trilogues. “I think that the meetings have shown that the agreements must be respected. The work of the groups enabled us to prevent this attempt to call the agreement into question. It totally failed. The EPP’s attitude will be entirely constructive. A very large majority of the group supported the report”, summarised Mr Benifei.
The text also provides for a ban on biometric identification systems using sensitive characteristics such as gender, ethnic origin or political orientation, and on predictive policing systems. Emotional recognition systems used in law enforcement, border management, the workplace and educational establishments would be banned, as would the untargeted capture of facial images from the internet or CCTV footage to create facial recognition databases.
Possible partial implementation before the elections
The question of how to classify high-risk AI systems was also an important part of the European Parliament’s internal negotiations. AI systems that can “seriously undermine” people’s health, safety and fundamental rights or the environment will be labelled as high-risk systems, as will AI systems used to influence voters or the outcome of ballots.
This provision comes against a backdrop of increasing disinformation and the forthcoming European elections. However, AI legislation will not yet be in force in 2024, although the co-rapporteurs are aiming to reach an agreement with the EU Council “before the end of the parliamentary term”.
Several avenues are nevertheless being explored to introduce safeguards before the elections. “I think we should discuss early implementation for foundation models and generative AI. This is a personal opinion, but in this specific area we need to see whether it is possible to speed up the entry into force”, commented Mr Benifei, recalling the importance of the Digital Services Act (DSA) (see EUROPE 13185/5) in this area.
Reinforced provisions for generative AIs
With regard to general-purpose AI and generative AI, system suppliers will have to “assess and mitigate potential risks”. Their models will have to be registered in the EU database before they can be put on the European market. Additional requirements exist for suppliers of generative AI systems, who will have to clearly mark AI-generated content so that it can be easily identified.
During the negotiations, the creative industry insisted on copyright protection. The text thus provides that detailed summaries of the copyright-protected data used to train AIs must be made public.
Finally, the report confirms the European Parliament’s ambition to strike a balance between user protection and innovation, with a number of provisions, such as the possibility for companies to use ‘regulatory sandboxes’. Companies will also be able to sign up, on a voluntary basis, to the future ‘AI pact’ (see EUROPE 13194/3), which should help them to be ready when the AI Act comes into force. Penalties of up to 7% of annual turnover could be imposed in the event of non-compliance, and there are also provisions for an AI system to be withdrawn from the market if necessary.
To see the document: https://aeur.eu/f/7HM (Original version in French by Thomas Mangin)