AI Safety: AI Safety Efforts Renew After Paris Summit Failure
After the failure of the Paris summit on theartificial intelligence Last February, the international community of security researchersIA is mobilizing again, hoping to contain the dangers linked to the rapid expansion of artificial intelligence. According to a report published Thursday, experts now agree on three priority areas to prevent existential risks such as the loss of human control or the availability of accessible biological weapons.
At a gathering in Singapore in late April, the tone was radically different from that in Paris. Max Tegmark, president of the Future of Life Institute and organizer of the conference, emphasized that "the majority of participants have regained their motivation, and hope is reborn." The scientific community now asserts that global convergence around the technical challenges of AI safety is underway. "The Americans and the Chinese agree that there is work to be done, and it must be done," says Yoshua Bengio, one of the principal researchers contributing to the report "Singapore Consensus on Global AI Safety Research Priorities."
The report highlights three key areas of research: assessing risks related to AI and its applications, developing safe and reliable AI by design, and monitoring deployed AI, with the ability to respond immediately if warning signals are detected. While the climate of fear remains palpable in the community, with apocalyptic scenarios evoking a total escape from humankind or the creation of bioweapons by poorly controlled AI, the scientific community emphasizes the potential impact of its work on guiding public policy.
The timing of the government summit on AI planned for India in 2026 should also reflect this optimistic tone, emphasizing the importance of research in positively influencing policy decisions. Tegmark reminds us that "the best way to obtain political will is through technical research. History shows that you don't need a nuclear winter for world leaders to take concrete action." The discussions in Asia, as impactful as the Paris summit, are helping to redefine the trajectory of AI security, emphasizing the need for a collective effort to prevent any future catastrophe.
Really?
Let's take a closer look at the key points of orientation:
- Risk assessment: (Black Box = Mystery) + (Alignment = They don't know how to do it) = Risk exposing 1000. Good luck with the assessment.
- Safe AI development: Safety is alignment. What percentage of researchers are aligned with the mass of AI engineers already? 1%? 10%? Good luck with safety.
- reliable from the design stage: ah! A slip of the tongue about the training paradigm, which is fundamentally incorrect and whose every step needs to be reviewed? There is hope, indeed, but unfortunately it is only on this point.
