A top AI safety researcher has issued a warning that society may be “sleepwalking” into a future where computers operate vital infrastructure before appropriate oversight is put in place, as the field’s rapid advancements in AI may surpass human capacity to put in place suitable safeguards.
The world “may not have time” to get ready for the safety threats posed by advanced AI systems, according to David Dalrymple, a program director at the UK’s Advanced Research and Invention Agency, who spoke to The Guardian. His cautions coincide with fresh statistics from the AI Security Institute in Britain showing that AI capabilities are developing at an exponential rate, with performance in some areas doubling every eight months.
The
Capabilities Progressing More Rapidly Than Controls
AI systems can currently accomplish apprentice-level cybersecurity jobs 50% of the time, up from less than 9% in late 2023, according to AISI’s first Frontier AI Trends Report, which was published in December. The first model that could perform expert-level cyber jobs, which normally require more than ten years of human experience, was evaluated by the institute.
The most alarming finding was that, in just two years, success rates on self-replication assessments rose from less than 5% to more than 60%, according to AISI. Although the institute states that models are “currently unlikely to succeed in real-world conditions,” the quick development highlights how quickly things are progressing.
“We should be worried about systems that can execute all human tasks more efficiently,” Dalrymple stated. “We risk being outcompeted across critical areas necessary to maintain control over our civilization, society, and planet.”
Economic Pressure Could Take Priority Over Safety
According to Dalrymple, the majority of economically significant tasks will be completed by machines in five years at a lower cost and with better quality than by people. He predicts that by the end of 2026, AI systems will automate the equivalent of a full day of research and development, which could lead to a feedback loop in which AI advances more quickly than policymakers can react.
The researcher stressed that governments shouldn’t presume that sophisticated systems will be dependable, especially in light of the financial pressure on businesses to launch products as soon as possible. “The science needed to guarantee safe behavior may not arrive in time given the economic pressure,” he stated. “So the next best thing that we can do, which we may be able to do in time, is to control and mitigate the downsides.”
The AISI report, which pointed out that protections are getting better, provided some consolation. Between some model generations that were introduced only six months apart, the period it took for skilled researchers to find universal jailbreaks grew by about 40 times. Nevertheless, the institute discovered flaws in each system that was examined.
Dalrymple, who is creating safeguards to protect AI’s application in vital infrastructure, such electricity networks, cautioned that “destabilization of security and economy” could result from technical advancements surpassing safety protocols. In order to better understand and regulate the behaviors of sophisticated AI systems before they become too strong to handle, he advocated for further technological study.