AI safety and security — topics currently under intense scrutiny following a series of incidents in which advanced AI models hacked real-world networks during testing activities — are not the core focus of the JRC’s testing activities, according to a different Commission official, who was granted anonymity to discuss sensitive material.
Open-weight models are those whose core components are publicly released, allowing anyone to run the model and study how it works. China’s approach to AI has leaned more towards open models than the United States. Some major U.S. tech and AI companies, including chips giant NVIDIA, are investing in open-weight models.
Forced to choose
The U.S. and China are locked in a race to dominate in the technology, with Europe largely sitting on the sidelines. Without a viable European frontier model, European private and public organizations may be left to choose between American and Chinese providers for advanced AI applications.
In a recent speech, European Commission President Ursula von der Leyen suggested Europe should not try to win the frontier AI race. A top German tech company told POLITICO this week it was dropping Europe’s leading AI lab Mistral in favor of open-source Chinese models.
The JRC has not entered into any “bespoke access” or licensing negotiations with the relevant Chinese developers, Berestecki said. The models are run on the center’s “secure, on-premises scientific infrastructure” and the Commission is not intending to use Chinese models for its own purposes, he added.
EU cyber agency ENISA — which has gained access to advanced models from top U.S. AI firms Anthropic, OpenAI and Google — is also experimenting with “a couple” of open-weight Chinese models, its executive director Juhan Lepassaar confirmed in an interview on Wednesday.
JRC and ENISA have been tasked with working together to set up an AI testing platform under a recent EU action plan on AI and cyber, the European Commission’s attempt to respond to the growing concerns after a series of incidents involving AI systems breaking out of testing environments and breaking into other organizations’ digital systems.