21st Century Business Herald reporter Feng Liange reports from Guangzhou
From 2021 to 2024, the EU's artificial intelligence law is gradually advancing.
Recently, the merged texts of two unofficial versions of the EU Artificial Intelligence Act (hereinafter referred to as the "Act") have been made public online. One of them is a comparative version of nearly 900 pages, while the other, with 258 pages, is a comprehensive document of the revised and basically finalized Bill. At the same time, the European Commission announced the decision to establish the European Artificial Intelligence Office (AI Office).
The interviewed experts pointed out that global artificial intelligence governance is racing, and AI governance is an important track. The accelerated pace of the European Union holds extraordinary importance for artificial governance on a global scale.
Granting regulatory exemptions for open source and research and development
What are the highlights of this public bill?
According to Zhu Yue, an assistant professor at the Law School of Tongji University and a researcher at the Shanghai Artificial Intelligence Social Governance Collaborative Innovation Center, the most noteworthy highlight of the text released this time is that it has left certain space for the development of the artificial intelligence technology industry in multiple aspects.
Clarifying the governance targets is a prerequisite for carrying out subsequent work. The bill has made certain changes to the core definition of artificial intelligence systems. It is mentioned that artificial intelligence systems are systems that may exhibit adaptability after deployment. The so-called adaptability refers to the autonomous learning ability that artificial intelligence systems may exhibit after deployment, allowing the system to undergo changes during use. This is likely to exclude artificial intelligence systems that rely solely on pre-set rules.
"Due to the fact that many algorithms may not necessarily have adaptability in daily understanding, this imposes a limitation on the scope of regulated artificial intelligence systems," he pointed out.
Secondly, the bill also provides exemptions for research and development (R&D) of artificial intelligence systems, stating that AI systems and models specifically developed and put into use for scientific research and development purposes, as well as their output content, must be excluded from the scope of the bill's regulation.
It is worth noting that the bill also expresses concerns about open source. "Open source is the foundation of the AI ecosystem, and it is very fundamental in the entire ecosystem. Open source is crucial in every aspect." Zhu Yue mentioned that the "Act" and the "Product Responsibility Directive" (hereinafter referred to as the "Directive"), which was also finalized in December, provide clear exemptions for open source AI.
For example, the Bill states that if the relevant tools, services, processes, or artificial intelligence components are provided under free and open source licenses, they should not be required to comply with the requirements for artificial intelligence value chain responsibility.
"Overall, the latest version of the bill provides a critical balance in terms of details," he said.
"The most dangerous moment has passed"
The implementation of the bill has not been smooth sailing.
Last November, France, Germany, and Italy had a disagreement over the EU's regulatory approach to basic models, requiring the use of model cards with lighter compliance obligations for regulatory purposes. In December, after 36 hours of marathon style trilateral talks, all parties have reached a political agreement on the bill.
However, the European Parliament will change in June this year. Is there still a risk ahead as the dust has not settled yet?
"The trilateral talks have invested a lot of time to form a basic consensus and push for the bill to be implemented as soon as possible. Therefore, in theory, the text will not undergo significant changes in the future," Zhu Yue said.
However, he also pointed out that some countries are not fully satisfied with the current provisions for public (real-time) facial recognition and general-purpose artificial intelligence. It may lead to another wave of discussion in the future council and parliament.
"However," he said, "the most dangerous moment has passed, and the probability of the bill not passing is very slim."
Today, the bill still needs five theoretical steps before it officially takes effect. Zhu Yue told 21 reporters that first, the relevant committees and plenary meetings of the European Parliament will vote, followed by the relevant committees and plenary meetings of the European Council. After completing two votes, the bill will be signed by the President of the European Parliament and the President of the European Council respectively, and then the constitution will be published. The bill will take effect twenty days after its publication.
After the bill is implemented, some key measures will begin to take effect at different points in time.
According to the currently available comprehensive text, the provisions of different parts of the bill will come into effect in stages after it takes effect as a whole. In terms of risk regulation for artificial intelligence systems, the ban on AI systems with unacceptable risks will apply for 6 months after taking effect; The obligations related to high-risk artificial intelligence systems will not apply until the 36th month after they take effect.
Why are there differences in applicable time nodes?
"Regarding the more obvious future risks, regulatory authorities hope to collect information or develop self-discipline codes of conduct as soon as possible, while relatively complex issues require longer time to test their effectiveness," Zhu Yue pointed out.
In his view, the core reason for this difference is that the bill involves too many stakeholders. "The regulatory authorities of the AI technology industry, as well as manufacturers, distributors, importers, operators, and deployers on the AI value chain, each have different perspectives and considerations. Dividing the effective dates and reaching compromises one by one may be a more reasonable decision."
In addition, the stock system is also a relatively obvious problem. After the implementation of the bill, some systems that have already been put into use may also need to be adjusted accordingly. How to integrate new laws with old rules for governance clearly requires some time for reflection and implementation.
Continuous improvement of governance structure
Although the bill has not yet been fully settled, European countries have decided to take a step forward.
Today, according to media reports, Germany will approve the Bill. Also this month, it was reported that the Netherlands will immediately implement the known parts of the Act, and the Dutch government hopes to make good use of generative artificial intelligence systems such as ChatGPT while taking risk prevention measures. A spokesperson for the Dutch Ministry of the Interior stated that although the bill has not been officially finalized, the Netherlands will immediately prioritize the parts it knows. "Enterprises will benefit from regulatory clarity," the spokesperson said.
As an AI governance "region", the EU's efforts go beyond the Bill.
Zhu Yue told 21 reporters that the European Commission is currently formulating a "strategic framework" to promote the development of artificial intelligence start-ups and the new generation of artificial intelligence. Its core concept will focus on "AI factories" - while also focusing on computing infrastructure such as supercomputers and data centers, focusing on building innovation communities and industrial upgrading, and providing policy support for chip and AI related investments.
Recently, the decision of the European Artificial Intelligence Office (AI Office) has also been released.
It is reported that this office will be part of the administrative structure of the Communications Network, Artificial Intelligence Content and Technology Bureau, and its functions mainly include promoting the implementation, monitoring and supervision of artificial intelligence systems, as well as artificial intelligence governance.
In addition, at the end of last year, tripartite negotiations between the European Commission, the European Council, and the European Parliament jointly reached a political agreement to amend the Directive, incorporating artificial intelligence into its scope of application. The current application status and development prospects of AI have been considered in the discussion of multiple important clauses such as negotiation definition, damage, defects, burden of proof, and mitigation of development risk defense.
"The implementation and promotion of the Artificial Intelligence Office and the Product Responsibility Directive mean that the EU is constantly accelerating the improvement of its AI governance structure." Zhu Yue pointed out that many developments indicate that AI governance at the EU level, with a broader significance, will not only discuss regulatory and regulatory issues, but also gradually form a dual track framework for innovation and regulation.