Can companies alone manage the risks of AI systems?

The Future of Life Institute argues that the risks of AI "cannot be managed by the companies alone," and that what is needed is "independent, binding oversight of AI." Its specific ask is that "countries should put pressure on companies to immediately limit the use of recursive self improvement until the research to make it safe has been done," a position it attributes to UN Human Rights Chief Türk. It pairs that limit with investment in "hardware verification technology" that could support "a future international AI deal involving both China and the U.S." The on-site inspectors proposed by Anthropic's Amodei and by OpenAI are, in its words, "excellent first steps, but we need more."

The argument runs as a chain. New incidents "of increasing severity" show an "urgent and vital need" for binding oversight; company measures and a proposed international institution fall short; so "private AI companies cannot be trusted to police themselves," and "government intervention is urgently required." The end point of the limit is set by research, not by a date or a capability threshold: it holds until the safety work is done. The statement also rules out one design, oversight that would create "a liability shield for these companies," and names hardware verification as the means that would make a cross-border deal checkable.

Against the library, the nearest governance note is How do we stop AI systems once they are already deployed?, which asks who can halt a deployed system. This statement asks for something earlier and broader: a limit on a practice while the safety research is still missing, not the stop of one system. Its key term also needs the distinction drawn in Are self-refinement and recursive self-improvement actually the same thing?. The excerpt does not define "recursive self improvement," so a limit on its use could cover bounded or open-ended forms, and the statement does not say which. What bottlenecks define the path from AGI to superintelligence? treats recursive improvement as one of four pathways; this statement treats it as a practice to restrain. The "incidents" the statement cites are unnamed, so Can defenders stop intrusions without knowing who sent them? is at most a candidate example; the excerpt does not connect the two.

The excerpt does not establish the premises the call depends on. It does not define recursive self improvement, name the escalating incidents, say which developers use the practice or how far, or describe how a limit would be monitored beyond the mention of hardware verification. It does not say what "safe" would require or who would judge that the research is done. Its weight rests on the officials and company proposals it cites, which the excerpt reports without the original text. At the strength the excerpt supports, the statement is a policy position with asserted premises. It establishes that its authors want a time-bound, research-conditioned limit backed by independent verification. Whether such a limit can be specified, and what it would cover, is left open.