Each political group of the European Parliament submitted a few hundred amendments to the upcoming AI Act, setting the tone for future discussions.
The amendments, thousands in total, prelude complex negotiations that are set to start before the summer, during which the co-rapporteurs Brando Benifei for the internal market committee (IMCO) and Dragoș Tudorache for the civil liberties committee (LIBE) will seek to reach a majority via compromises.
According to a European Parliament official, the MEPs in the two leading committees are almost equally split around the centre-right and centre-left axis, and on the most controversial points, it will be a ‘calculation game’.
Definition
One of the most controversial topics is the definition of artificial intelligence (AI) itself.
MEP Benifei is proposing a broad definition and deleting the list of AI techniques and approaches in Annex I to make the regulation future-proof.
By contrast, the centre-right European People’s Party insists on the definition agreed upon at the OECD level. The EPP also introduced a definition of machine learning as the ability to find patterns without being explicitly programmed for a specific task.
The two co-rapporteurs maintain their opposition to curbing an exception for general-purpose AI. In contrast, the EPP proposed to ease the burden of obligations for these providers, introducing different requirements for new, former, and original providers.
Scope
Liberal Tudorache introduced a new article to capture AI applications in the metaverse in the scope of the regulation, including a reference to blockchain-backed currencies and NFTs. He also proposed applying the rules to providers not located or operating in the EU under certain circumstances.
Prohibited practices
Green MEPs made major proposals in terms of prohibited practices, extending this category to biometric categorisation, emotion recognition, and any automated monitoring of human behaviour. These include recommender systems that systematically suggest disinformation and illegal content, uses for law enforcement, migration, work and education.
Biometric identification and recognition
In a major change of position, liberal Tudorache joined the social democrats and greens in advocating for a complete ban on biometric recognition, eliminating the exceptions included in the original proposal.
On biometric identification, Tudorache and the EPP both propose to exclude the consensual authentication to access a service, device or premises from the definition.
The Green group introduced a ban on private biometric databases based on scraping information from the internet, such as the controversial firm Clearview AI.
High-risk systems
The centre-right put forth a new regime changing the automatic classification for systems falling in the list of sectors mentioned in Annex III in a list of ‘critical use cases’.
Based on these uses, AI providers would have to self-assess if their systems pose significant risks to health, safety and fundamental rights.
Moreover, the EPP proposed that the obligations for high-risk applications might be partially or completely removed if, by fulfilling the regulation’s risk-mitigating measures, the systems have sufficiently mitigated the risk with operational countermeasures or built-in features.
Conservative lawmakers also excluded systems designed to assess creditworthiness from the high-risk list. In contrast, green MEPs extended this category to media recommender systems, algorithms used in the health insurance processes, payments and debt collection.
The green group made much stricter environmental requirements and increased the involvement of third parties and the notified bodies. Additionally, fundamental rights impact assessments have been made mandatory for all providers and deployers.
Legal clarity
The EPP proposed the creation of a new article defining trustworthiness by introducing a set of principles on how the concept should be applied to AI in terms of requirements and technical standards.
Moreover, to better define how the concept of accuracy, reliability, robustness and cybersecurity should be applied, conservative lawmakers introduced provisions requiring national metrology and benchmarking authorities to provide non-binding guidance.
Governance
The EPP followed the French Presidency’s proposal to involve stakeholders in the work European Artificial Intelligence Board and proposed to give the body more autonomy in setting its own agenda.
The Greens want the European Data Protection Supervisor to provide the secretariat for the Board. The EDPS would act as the supervisory authority for large companies whose breaches have a ‘community dimension’ as per the EU’s Merger Regulation criteria.
Lawmakers from both sides of the aisle seem to agree on giving more investigatory power to the supervisory authority, with the greens particularly ambitious in terms of corrective powers.
Advertising and dark patterns
A separate amendment was presented by MEP Maria-Manuel Leitão-Marques, with the support of the Tracking Free Ad Coalition, to include AI systems used to deliver online advertising to the list of high-risk systems.
The Green group added a paragraph providing that the transparency requirement for automated decision-making should contribute to addressing power imbalances in the digital environments and countering dark patterns.
Fines
Both liberal and conservative MEPs are proposing a general lowering of the fines, with the EPP in particular including a carve-out for SMEs and adding factors to consider such as intent, negligence and cooperation when the authorities calculate the fine.
By contrast, centre-left Benifei is pushing for an overall increase of the sanctions and for removing size and market share consideration from the criteria authorities would have to consider when imposing a penalty.
[Edited by Alice Taylor]