Scientists demand international AI oversight

More than 400 scientists are urging the creation of an international body to oversee the safety of artificial intelligence, according to an interview with Sciences et Avenir by Laurent Daudet, executive director of AI at the École Normale Supérieure (ENS). The group includes prominent figures in the field, such as Geoffrey Hinton and Yoshua Bengio, who are considered pioneers in artificial intelligence.

Major AI companies support regulation

This call for oversight aligns with recommendations from major U.S. companies in the AI sector, including Anthropic and OpenAI. These companies, which developed the AI models Claude and ChatGPT respectively, have also expressed the need for urgent regulation of AI technologies.

Closed AI models and testing

A common feature of AI systems like Claude and ChatGPT is that they are "closed" models, meaning their internal parameters are not accessible to users. Despite this, their behavior can still be tested through thousands of queries to evaluate whether they could produce harmful instructions or bypass restrictions.

Proposed AI certification process

An independent regulatory body could assess AI models by submitting numerous queries to test their responses, much like testing a car’s safety without needing full knowledge of its internal design. Certification would focus on observable behaviors rather than the full internal workings of the AI.

Open AI models and their implications

Some AI models, like DeepSeek from China, are described as "open," suggesting their internal parameters may be accessible. However, the specific implications of this openness were not fully explained in the interview.