In response to the global artificial intelligence industry supervision,一刻资讯 EU legislators will set off a new round of "Brussels".
"Internet Legal Review" today compiled the latest article by Natasha Lomas, a senior reporter of TechCrunch in Europe, for Chinese readers to understand the latest legislative rules for artificial intelligence supervision, as well as unproductive matters.
After nearly three days of marathon talks, the European Union Legislator -European Parliament, Council and the European Commission -on December 9, 2023 reached a political agreement on the first comprehensive regulatory rules of artificial intelligence, regulating artificial intelligence in the standard of artificial intelligence.Use in tools such as biometrics and monitoring.They will discuss the details of the final legislation in the next few weeks, and the legislation is expected to take effect early next year and be implemented in 2026.
EU Chairman Ulsla von Delin(Ursula von der Leyen)Praise this political agreement is "the world's first".When she took office at the end of 2019, she will formulate a key priority of promoting "trustworthy" artificial intelligence as a key priority.
Before the final text compilation and disclosure, all the details of the agreement reached may take a few weeks to fully confirm.Comb with public documents just released by the European Parliament.
Artificial intelligence prohibition matters
Fully prohibit the use of artificial intelligence in the following areas:
The EU law enforcement department's use of long -range biometric technology in public places has not been completely banned, but the European Parliament said that the negotiating representatives have reached an agreement on a series of guarantee measures and a few exceptions to limit the use of facial recognition and other technologies.This includes a criminal list that needs to be obtained in advance, and the use is limited to "strict definition".The real -time use of this invasive artificial intelligence technology will be limited by time and place, and can only be used in the following goals:
The EU Council emphasizes that the Artificial Intelligence Act will not be applicable to systems specializing in military or defense, as well as artificial intelligence systems that are not used only for research and innovative purposes, or people who use artificial intelligence for non -professional reasons.
Private social groups are suspicious of this. They are worried that restrictions on the use of biometric technology for national institutions will not be enough to protect human rights.
The rules of "high risk" AI and general AI
"High -risk" artificial intelligence system includes artificial intelligence that "has great potential harm to health, basic rights, environment, democracy, and the rule of law.
The European Parliament Members have successfully incorporated the compulsory basic rights impact assessment of the basic rights of the European Parliament, which is also suitable for insurance and banking.The artificial intelligence system used to affect the election results and voters' behavior is also classified as high -risk. At the same time, citizens will have the right to complain to the artificial intelligence system and obtain an interpretation of decision -making based on high -risk artificial intelligence systems.
The meeting also agreed to apply the "two -layer" guardrail system to the "general" artificial intelligence system, such as the so -called basic model that supports the growth of virus -type growth of virus -type virus such as ChatGPT.
Artificial Intelligence General Model(GPAI)The basic model will be constrained by the requirements of transparency, such as drafting technical documents, compliance with the EU copyright law, and the detailed summary of the content used in the training of relevant algorithms.For "high influence" GPAI with so -called "system risks"(Defined as the cumulative calculation of training measured in the floating -point operation is greater than 10^25)There is a stricter obligation:
The basic model of commercialization/GPAI is facing the supervision of the bill, but the R & D is not the scope of the law -according to today's statement, the supervision requirements of a complete open source model will be lighter than the closed source model.The agreement also promotes the establishment of regulatory sandboxes and real -world tests of the national authorities to support startups and SMEs develop and train artificial intelligence before the artificial intelligence release market.
Punishment and effectiveness
According to the regulations of the European Parliament, the punishment for illegal acts may lead to 7%of 35 million euros or 7%to 7.5 million euros or 1.5%of the turnover, depending on the violations and the size of the company.
Violations of prohibited artificial intelligence applications will apply higher sanctions(7%)And the behavior of providing incorrect information will be fined 1.5%.In addition, the temporary agreement allows the administrative fines of SMEs and startups to set the "more competent upper limit" on the administrative fines of SMEs and startups.Therefore, the infringement punishment faced by artificial intelligence startups seems to be smaller than artificial intelligence giants.
The agreement also allows 6 months before the law of the law is effective in the post -division stages of the law. It has 6 months before the rules of the use of use cases.Therefore, the entire effectiveness of the EU "Artificial Intelligence Act" may not be felt until 2026.
State Secretary of the State Secretary of the Digital and Artificial Intelligence State Secretary Kamame Atgas(Carme Artigas)Leading the council's negotiations on the document, she predicts that the law will support them "legal certainty and technical certainty" by giving European developers, startups and future scale expansion.
Is the road ahead unblocked?
Although the European Union expressed appreciation on the agreement reached by the "world No. 1" artificial intelligence rules, this is not the end of the EU legislation, because there are still some formal steps that need to be adopted -especially the final text will vote in parliament and council.
EU "Artificial Intelligence Act" to be discussed unspoken:
However, in view of what(Even)There are so many differences and disputes in regulating artificial intelligence. This political agreement has eliminated the biggest obstacle, and the road to the EU Artificial Intelligence Act in the next few months looks clear.
The European Commission undoubtedly showed confidence.The work of executing the agreement will start immediately -setting up an artificial intelligence office in the EU administrative department, which will be responsible for coordinating with member states supervision agencies that need to implement the rules of artificial intelligence companies;Promote the formulation of standards and test practice.The EU will appoint a scientific group composed of independent experts and provide suggestions to the Artificial Intelligence Office on the GPAI model.
In recent weeks, Mistral, France and France, took the lead in opposing the obligation to incorporate general artificial intelligence layered rules into the artificial intelligence package plan. The company has been lobbying to completely stripped the basic model/ GPAI.If the agreement agreed by the President of Spain does contain some obligations of GPAI and basic models.Therefore, this is not a "completely stripped" that Mistral and lobbyers have been pushing.
In response to the relevant political agreement, the French Digital Minister's Office said that he would carefully analyze the compromise plan and "maintain its strategic autonomy."