Published on Dec 10 2023 at 12:16
Intense, sometimes heated, even endless debates… But the world’s first new rules to control artificial intelligence (AI). Detailed review of the most significant EU provisions of this new legislation.
· High risk systems
The general approach of the law consists of imposing obligations on systems deemed “high risk”, therefore potentially causing harm to health, safety, fundamental rights, democracy, etc. For example, AI systems used to influence the outcome of elections.
The higher the risk, the stronger the bonds. They require an impact analysis on fundamental rights for any AI model before being placed on the European market.
But also obligations of transparency on the use made of these AI systems, or even explanations to citizens on those having an impact on their rights. Actors with emotion recognition systems will, for example, have to signal them to users.
· Generative AI
The art and manner in which the very fashionable generative AI must be regulated have generated so much debate that at one point, the entire text was almost called into question…
Ultimately, the pear was split in two between those who, like France, advocated self-regulation, and others (European Parliament, many scientists, etc.) who wanted strict obligations.
A double-speed system provides for stronger obligations, even if they remain limited, for the most powerful models. The others will only have transparency obligations, such as comply with EU copyright law For example.
Concretely, potentially systemic actors, such as the models underlying ChatGPT from OpenAI or Bard from Google , will above all have to ess, monitor and mitigate the risks of their systems themselves, or even report serious incidents to the Commission. Measures supplemented by a voluntary code of conduct, “until harmonized European standards are published”, specifies the Parliament.
In a tweet saturday , Yann LeCun, one of the founding fathers of AI now working for Meta, welcomes the fact that the legislation provides for exemptions for so-called “open source” models. This favors those of Meta but also European start-ups like Mistral and Aleph Alpha.
· Facial recognition
The text formulates several prohibitions including the very controversial real-time facial recognition, these cameras which scan faces.
But it poses some exemptions. Many EU countries, including France, wanted to be able to use it in matters of national security. This led to fierce exchanges between legislators – Parliament notably wanted a total ban.
Ultimately, it is authorized in the case of searching for victims, preventing a terrorist threat or locating and identifying people suspected of crimes.
Law enforcement also benefits from other exemptions to use remote biometric identification of people in public places.
The agreement, however, bans the ultra-controversial biometric analysis which categorizes people according to sensitive characteristics, such as political or religious beliefs, orientation or race.