The EU is drafting the world’s first AI regulation – you should know that

The institutions of the European Union are not known for their speed. But today, Wednesday, the EU Parliament actually entered uncharted territory: MEPs voted in favor of a draft that aims to regulate the use of artificial intelligence (AI) across the EU. If the law comes into effect, the EU would become the first continent in the world to regulate AI.

The speaker of the parliament, Roberta Metsola, spoke of “legislation that will set the world standard for years to come”. But what exactly does the so-called AI law say – and what problems are there with the specifications? Here’s an overview:

How should artificial intelligence be regulated?

In general, the design provides for the classification of artificial intelligence into different risk areas. Depending on how high the risk is, different requirements apply:

Artificial intelligences that pose an unacceptable risk will be banned. This includes all sorts of social scoring technologies. These have been used in China for years to rate and classify people based on their social behavior.

The real-time use of biometric data, such as AI-supported history recognition software, is also not allowed. Records, on the other hand, may be used, but only with law enforcement and with a judge’s approval.

What is considered a risk area?

The highest security level is the risk area. Think, for example, of AIs that are used in critical infrastructure. This area also includes AIs that can be used to influence elections or used in major social networks with more than 45 million users, for example to recommend new content.

There must be strict controls here: constant human control must be guaranteed, as well as special security measures. In addition, it must be documented with which datasets the AI ​​has been “trained”, i.e. which raw data has been entered into the system.

What about highlighting AI-generated content?

For low-risk applications such as chatbots it should suffice to point out to users that they are using an AI. Because probably many people are no longer aware of the use of AI in daily life.

At the latest with the emergence of ChatGPT, which brought the functions of AI closer to a very wide audience for the first time, it has become clear that many AIs generally cannot limit themselves to risk areas because they have an extremely broad scope of application.

For such a Foundation models Special rules must apply: if an artificial intelligence produces images or texts, for example, it must always be clear that the content has not been created by human hands. The AI ​​must also be designed in such a way that it is not possible to generate illegal content.

What are the weaknesses of the law?

The basic idea of ​​the proposal that AI should be largely monitored by humans and broken down into risk areas raises several questions:

The regulation is also widely criticized because it is not only the developer side that fears over-regulation. In particular, the “Foundation Models” requirements can incur huge costs, making it significantly more difficult to establish and operate AI startups on European soil. This allowed developers to continue looking for luck in Asia or the US.

At the same time, it is to be feared that the requirements will favor the big tech giants such as Microsoft, Apple or Google, whose huge financial capital makes it easier to meet EU requirements.

What happens after the vote?

The law has not yet been passed. The next step is the so-called trilogue. The three EU bodies Parliament, Commission and Council agree on a common position. Negotiations on this should start on Wednesday evening. The aim is to agree on a common law by the end of the year.

What were the positions in the EU Parliament?

As early as 2021, the European Commission presented a first proposal for the regulation of AI. The document was subsequently tightened in some places by EU parliamentarians and approved in its current form in May in the Committee on Justice and the Internal Market. Also new is the regulation for “Fundamental Models”.

The leading groups in the EU parliament (EPP, Social Democrats, Greens and Liberals) originally agreed to submit a joint proposal. However, the conservative EPP group has long wanted a less strict interpretation of the use of biometric data and therefore came up with a proposal that would allow real-time use in certain situations for law enforcement.

Shortly before the vote in Strasbourg, it became clear that other groups were also not adhering to the agreement: in total, more than 30 amendments to the compromise were tabled by the committees, including from the ranks of the Social Democrats and the Greens. In the end, however, Parliament voted in favor of the original proposal.

Sources

(t-online/dsc)