According to recent reporting by The Decoder, Redwood Research chief scientist Ryan Greenblatt has estimated the risk of an artificial intelligence takeover at 50 to 60 percent if current development trajectories continue. Greenblatt points to race dynamics inside major organizations like Anthropic and OpenAI as primary drivers of the ongoing arms race. During discussions, author Sam Harris noted a stark contrast between this high-risk environment and historical precedents, observing that Manhattan Project scientists would have halted their work at a 10 percent risk level. Greenblatt ultimately suggests that avoiding worst-case outcomes will require an international agreement.
The Mechanics of Competitive Pressure
The core issue highlighted by safety researchers is a collective action problem within the artificial intelligence sector. Major laboratories consistently view their own internal practices as adequately responsible, even as the aggregate effect of their competition accelerates development timelines. When every organization believes it is the uniquely safe steward of powerful technology, competitive pressure overrides caution. This dynamic mirrors classical arms races, where mutual distrust and the fear of falling behind compel participants to accept escalating probabilities of catastrophic outcomes.
What Remains Unclear
The reporting did not specify the exact mechanisms or terms of the international agreement Greenblatt believes is necessary, nor did it outline which geopolitical bodies would enforce such a framework. Furthermore, the announcements from Anthropic and OpenAI regarding how they plan to reconcile competitive pressures with stated safety commitments were not detailed. Observers will need to watch whether diplomatic channels can establish binding limits on artificial intelligence development before thresholds of severe risk are reached.
As laboratories push forward, the tension between commercial competition and existential risk management remains unresolved.
Lead image: Sora Shimazaki / Pexels
