Struggle for AI Act: EU Parliament targets ChatGPT

With ChatGPT hat artificial intelligence (AI) get a face for the first time. Average users can easily ask questions to ChatGPT, view artificially generated images from programs such as Midjourney can be created or create videos from text with just a few clicks.

Behind these programs, which have moved into the mainstream of society, is so-called “generative AI” – a collective term for artificial intelligence that produces original content without human intervention. Actually, the EU institutions have been working for 2021 in it, within the framework of AI Act Establish guidelines for AI applications. However, the EU did not have generative AI models, such as those behind ChatGPT, which are not tied to a specific purpose and can have a wide variety of applications, on their radar. Until recently there was no trace of ChatGPT and Co in the drafts.

Now that comes EU-Parliament however, in the aisles. The responsible members of parliament currently have a draft on the table that also includes generative “General Purpose AI” should regulate. The draft will be available on Thursday approved by the parliamentary committee. An important vote, since essential content of the AI ​​Act is defined here. And there is still not agreement on all points.

“Have to document training data”

“(The AI ​​Act) was not an easy task because it is the first time in the world that AI has been regulated in this way,” says Brandon Benifei, lead MEP, on the AI ​​Act in a press briefing on Wednesday. “We have tried to create a framework that is flexible enough for future innovations, but at the same time protects citizens”.

Parliament has agreed on a “special regime” for general-purpose AI. Companies that develop generative AI will have to use their training data in the future document. This is primarily intended to enforce copyrights enable, since AI applications such as Dall-E or ChatGPT train on texts and images that can represent intellectual property.

“You have to from now on document thoroughly and make transparent what material was used to train an AI,” says MP Dragos Tudorache firmly. This is the only way those affected are able to assert a violation of their copyrights. The content that generative AI produces must also not violate the law. For example, fake news or deep fakes should be stopped. If manufacturers of generative AI do not default, there is a risk of one fine.

High-risk category is intended to regulate sensitive areas

Contrary to what was originally intended, however, generative AI should not be used as a standard high-risk application get ranked. Stricter rules apply to these applications. For example, they may not be used by everyone and must also have a kind impact assessment, conducted by the competent authorities. According to the current draft, high-risk AIs include systems that are used in elections to count votes. Or those that are used in insurance or lending.

The subdivision of AI applications into 4 categories remains, as announced in advance: In addition to the high risk, there are also the levels of minimal risk, low risk and unacceptable risk. Unacceptable and therefore prohibited are, inter alia, the exploitation of particularly vulnerable persons, including Kinder. Also the Real-time remote identificationn, so the mass surveillance per biometric facial recognitionis inadmissible according to the parliamentary draft.

Real-time facial recognition still pending

However, it is precisely a ban on such AI systems not set in stone. MP Tudorache emphasizes vehemently: “In the EU you will not be able to use biometric facial recognition in real time. Point”. And member Benifei also describes the technology as a “security policy nightmare”.

For some MPs Conservative Group (EPP) However, this position in the AI ​​Act draft should be difficult to digest. Because she traditionally advocates the use of biometric facial recognition in criminal prosecution and could therefore thwart her on this point. And that too Ratthe represents the governments of the EU member states, could be more positive about the technology.

“We see several points in which our positions are similar,” Benifei is optimistic about the further legislative process. As soon as the MEPs have agreed on a common position, the trilogue negotiations between Parliament, Council and Commission to start. This is likely to be the case in autumn. The AI ​​Act is then expected to 2025 come into effect.