The comprehensive guidelines governing AI in Europe could set a precedent for other regions.
After a series of intense negotiations, lawmakers in Brussels have reached a “provisional agreement” on the European Union’s proposed Artificial Intelligence Act (AI Act) this week. This act is poised to become the world’s second detailed set of regulations overseeing AI and may establish a standard for other jurisdictions considering similar frameworks.
The agreement includes provisions for “high-impact” general-purpose AI (GPAI) systems that must adhere to specific criteria such as risk assessments, adversarial testing, incident reporting, and more. It also requires transparency from such systems, including the creation of technical documentation and “detailed reports on the data used for training” – a practice that companies like OpenAI, the manufacturer of ChatGPT, have previously resisted.
Furthermore, individuals will have the right to lodge complaints about AI systems and receive explanations regarding decisions made by “high-risk” systems that affect their rights.
While the press release did not delve into the operational specifics or the exact measures outlined, it did mention the establishment of a framework for imposing sanctions on companies that violate the regulations. The fines will vary depending on the violation and the size of the company, ranging from \(35 million or 7% of global turnover to \)7.5 million or 1.5% of worldwide revenue.
The guidelines also outline several instances where the use of AI is prohibited, such as extracting visual data from CCTV footage, making decisions based on sensitive attributes like race, sexual orientation, religion, or political beliefs, implementing emotion recognition in educational or professional settings, or developing social credit systems. Additionally, AI systems that manipulate human behavior to undermine their autonomy or exploit individuals’ vulnerabilities are strictly prohibited. The regulations also provide safeguards and exemptions for law enforcement’s use of AI systems in real-time monitoring or data searches.
It is anticipated that a final agreement will be reached by the end of the year. However, the enforcement of the law is not expected until at least 2025.
The initial draft of the EU’s AI Act was introduced in 2021 with the aim of defining AI, establishing regulatory principles for Artificial Intelligence across EU member states. The emergence of rapidly evolving AI technologies like ChatGPT and Stable Diffusion necessitated several revisions to the policy.
Further negotiations are required following the provisional agreement, including votes by the Internal Market and Civil Liberties committees of Parliament.
Discussions surrounding the regulations governing live biometric surveillance (such as facial recognition) and foundational AI models like OpenAI’s ChatGPT have been particularly contentious. The announcement of the agreement was delayed as these matters were reportedly still under discussion until the recent announcement.
While some governments have advocated for exemptions for military, law enforcement, and national security applications, EU legislators have advocated for a complete prohibition on the use of AI for biometric surveillance. Recent proposals from France, Germany, and Italy to allow developers of relational AI models to self-regulate have also contributed to the ongoing debates.
