The European Parliament votes on a “historic” artificial intelligence law without resolving key dilemmas | Technology


This Wednesday the European Parliament will vote on the European Regulation on Artificial Intelligence (AI), agreed on December 8 by the Commission, the Council and the European Parliament itself after a marathon 38-hour negotiation session, the longest in the European Parliament. history of the EU. The final draft of the regulations, which has not been revealed until this week, confirms what could already be seen in December, when the provisional agreement was closed, which also still requires the final ratification of the Member States before it can come into force. : The regulation outlines what can and cannot be done, but does not specify how key issues should be resolved, such as the defense of copyright.

The rule is considered the first AI regulation that protects basic rights and freedoms (China launched its own law last summer). In the last debate before the vote, the Internal Market Commissioner, Thierry Breton, promoter of the regulations, defended it on Tuesday in Strasbourg as the “first AI regulation in the world balanced and intended to protect against the excesses of misuse of AI, but at the same time promoting innovation.” The regulations, Breton insisted when requesting majority support from the European Parliament, represent a “time-tested historic commitment.”

The regulation establishes different requirements and obligations for AI applications depending on the risks posed by their use. The most innocuous ones, such as spam filters or duplicate text detectors, can be used without any restrictions. They are called a limited risk system and the only requirement placed on providers is that they inform users that they are using an AI tool.

The bulk of the negotiation consisted of determining which ones are of unacceptable risk and, therefore, are totally prohibited; and which are high risk, which implies permanent supervision. The first category includes systems “that transcend a person’s consciousness or deliberately manipulative techniques”, those that exploit their vulnerabilities or those that infer people’s emotions, race or political opinions.

Included in the high-risk heading are remote biometric identification systems, which a large sector of Parliament wanted to strictly prohibit, biometric categorization systems or emotion recognition. Also systems that affect the security of critical infrastructure and those related to education (behavioral evaluation, admission and examination systems), employment (personnel selection) and the provision of essential public services, law enforcement or migration management.

The document, of 460 pages, does not incorporate relevant surprises, although it confirms that the lack of definition surrounding some of the most sensitive issues of the regulations remains unresolved. “It is a rather brief regulation. As a legal instrument, it requires more explanation. There are very ambivalent definitions, some of them coterminous with those of other regulations,” says Lorena Jaume-Palasí, expert in ethics and philosophy of law applied to technology and advisor to the European Parliament on issues related to AI. “It seems likely that, just as has happened with the General Data Protection Regulation (GDPR), we will need jurisprudence to know with what methodology all of this will be implemented.”

Biometric identification: in exceptional cases

One of the most controversial issues, which took up the most hours of negotiations, is the regulation of the so-called real-time remote biometric identification systems. This category would include, for example, the extensive network of facial recognition cameras spread throughout large Chinese cities that allow anyone to be located in a few minutes. The first drafts of the European AI regulations were very restrictive with this technology.

In the latest version of the document, however, a series of cases are established in which they may be used, always under judicial authorization: “The search for certain victims of a crime, including missing persons; certain threats to the life or physical safety of natural persons or threats of a terrorist attack; and the location or identification of the perpetrators or suspects of (a list of 32) criminal offenses.”

The final text also includes an exception about exceptions: “In duly justified cases of urgency, such systems may begin to be used without registration in the EU database, provided that such registration is carried out without delay.” improper. (…) Member States shall notify these rules to the Commission no later than 30 days after their adoption.” That is, the authorities can bypass the regulations and act, even if it is later declared that the use of the system has been inappropriate.

“The more exceptions a ban has, the less harsh it is. What we are seeing is that technologies that should be questioned are being naturalized,” says Jaume-Palasí.

During the last debate in the European Parliament, those mainly responsible for negotiating the regulations have, however, defended the AI ​​Regulation as an instrument that will allow citizens to use these new technologies without fear of their fundamental rights being violated.

Copyright: protection, but not specified

Another of the issues of the regulation that has attracted the most attention is how copyright is protected, enshrined in the EU through 13 directives and two regulations. In the US there are a series of class action lawsuits filed against large platforms for using, without permission, creations of individuals to train AI models; The result of these processes can mark the future of generative AI (which generates texts, images, videos or music from instructions provided by the user).

The European AI Regulation repeats throughout its articles that all systems must guarantee compliance with copyright, but it does not say how, beyond the fact that it will be the responsibility of the AI ​​Office, the body that will supervise compliance. of the regulations. According to Breton, this office will begin operating as soon as possible once the regulation comes into force and will have “the necessary talent to support the implementation of the law and to serve as a center of global expertise in AI.”

Regarding the copyright of the content used to train algorithms already in operation, “the regulation is quite guaranteeing and will provide a lot of legal security to citizens. More than a legal framework, it is a declaration of democratic principles and values,” emphasizes Iban García del Blanco, socialist MEP and negotiator of the AI ​​law. ”In some cases, the Commission will have to develop more than 20 acts derived from these provisions to make them concrete. For example, in terms of copyright: it will prepare a series of standard proposals so that companies know in what sense they will have to make these transparency declarations about the content they have used in the training of the models,” he explains.

This does not prevent, says García del Blanco, that in the future there will be a series of demands and that criteria will be established through jurisprudence. Jaume-Palasí recalls that this is what happened with the data protection regulation: “The first rulings are coming out now, five years after it came into force.”

You can follow The USA Print in Facebook and x or sign up here to receive our weekly newsletter.