The European Union reached a deal Friday on the world’s first regulations for artificial intelligence (AI), paving the way for legal oversight of the popular technology used in services like OpenAI’s ChatGPT.
The European Parliament and its 27 member countries had to overcome major differences on a number of sticking points, including generative AI and facial recognition. There were also disagreements on how and whether to regulate the foundation models of chatbot software such as ChatGPT and Google’s Bard, according to the Associated Press.
In addition, EU lawmakers weighed full bans on AI-powered facial recognition systems due to concerns about privacy, putting them at odds with member countries who want to use the systems for law enforcement. As a result, talks were paused Thursday after 22 hours of negotiation left them at a stalemate.
The talks began Wednesday in Brussels and resumed on Friday morning.
Europe took the lead in the race to develop rules around AI in 2021, which has spurred concerns about the risks the advanced technology could pose to privacy, jobs and copyright protection.
The deal will still need to be voted on by the European Parliament early next year, but Italian lawmaker Brando Benifei, who helped lead negotiations, told AP that is seen as a formality.
The deal will require foundation models, including ChatGPT and other general purpose AI systems, to comply with transparency obligations before they are on the market, according to Reuters.
Obligations include:
- Drawing up technical documentation
- Complying with EU copyright law
- Disseminating detailed summaries about training content.
The deal also requires foundation models with a high impact and systemic risk to do more, including:
- Conducting model evaluations
- Assessing and mitigating systemic risks
- Conducting adversarial training
- Guaranteeing cybersecurity
- Reporting on energy efficiency
- Reporting to the European Commission on serious incidents
As far as surveillance, governments can use real-time biometric surveillance, but only in public spaces for victims of certain crimes, to prevent present or foreseeable threats, including terrorist attacks and to conduct searches on those suspected of the most serious of crimes, per Reuters.
The agreement also bans cognitive behavioral manipulation, scrapping of facial images from the internet or CCTV footage and scoring and biometric systems that infer religious, political, philosophical beliefs, along with race and sexual orientation.