Here’s How an AI Slowdown Could Actually Be Enforced

The potential for a slowdown or pause in artificial intelligence (AI) development has gained significant attention in recent weeks. As worries about the risks associated with increasingly powerful AI systems mount, leading figures from the industry have called for such measures to be taken. However, experts argue that implementing these restrictions will require a concerted effort from various stakeholders, including outside researchers and authorities.

One proposed solution involves granting third-party evaluators greater access to AI models. These evaluators would test the models' capabilities and "red team" them by attempting to elicit misbehavior in secure environments. Some experts believe that rigorous inspections and audits could effectively pause the development of frontier AI. Nonetheless, these doomsayers argue that the evaluations must be independent and scientifically rigorous, suggesting that the involvement of agencies like the FBI or NSA may be necessary.

Another approach gaining traction includes imposing limitations on the compute required for AI development. The existing 2023 Biden-era AI executive order already requires companies to report training runs above a certain compute threshold. Researchers propose that cloud providers could play a crucial role in future efforts by providing visibility into major AI training runs through tracking cloud usage, GPU utilization, network traffic, and power consumption.

Some even suggest modifying the chips themselves to log and verify compute runs, allowing for periodic inspection. However, experts stress that more research is needed to fully understand AI models and develop effective evaluation techniques. While some AI companies have proposed using their own evaluators, outside researchers argue that these assessments must be more rigorous and independent to ensure AI remains safe and aligned with human values.