International body for advanced AI: What scientists are asking for and why
Follow Us
Approximately 400 scientists from 34 countries, including leading figures in artificial intelligence such as Geoffrey Hinton and Yoshua Bengio, have signed a letter asking for the formation of an international organization responsible for monitoring containment and reporting of advanced AI systems capable of conducting cyber operations. The letter was released on Tuesday and claims that industry pledges are not enough.
SurveyAlso read: Nothing smart glasses leak hints at two models, one with AI and a camera: Report
What the scientists want
A common theme is the establishment of an ongoing international body with a well-defined mandate. As Reuters reports, it will establish the standards of containment in the industry and keep a registry of security breaches, which, according to the scientists, must eventually be released into the public domain.
Containment is the crucial word here. In their letter, the authors argue that AI models at the frontier must first of all be air-gapped, i.e., disconnected from any other systems, and then independently overseen. The reasoning is straightforward – a model cannot breach another system if it cannot access it. The letter does not call for a new body to be established. The organizations developing or testing cyber-capable models must implement high-assurance containment standards right now, the signatories insist.
Why now and where governments stand
The letter was written after several cases were identified by researchers as demonstrating problems with the process of testing and containment of such advanced AI systems. It includes situations when AI models escaped from sandboxing, which is the process used to separate the system from all others during its testing process.
Also read: Workday CEO Aneel Bhusri calls Oracle, SAP as AI ‘dinosaurs’
It is also timely politically speaking because the letter was sent one month after five firms including Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a White House accord promising themselves voluntarily safety mechanisms that do not have any means of enforcement. The message of the letter is that promises without enforcement are not enough.
Policymakers are currently arguing over the methods that can be used to supervise such systems, and the answers are moving in various directions. EU commissioner for innovation and digital policy Henna Virkkunen said to Reuters on Friday that AI Act is more than enough to deal with rogue AI systems. And then on Sunday, the Trump administration removed its embargo on US labs sending Britain’s AI Security Institute access to their models for safety testing. It creates opportunities for international tests, but still, a lot needs to be done to create an international institution.
Not everyone is on board
This letter is not fully backed by all prominent individuals in the field. The Turing Award was received by Geoffrey Hinton and Yoshua Bengio jointly with Yann LeCun, one of the directors of Meta, who also did not sign the letter. It seems that according to him, the recent security threats mostly stem from the mistakes of engineering, not from any existential threats posed by AI. Hence, the main question which forms the basis of the debate emerges here: whether those threats need to be addressed by establishing some international institution or simply through better engineering. Afterall, an open letter is not an institution.
Also read: Apple Home Collection cameras: No camera subscription but you still pay iCloud+
