Every AI lab thinks it's the responsible one, and safety researcher Ryan Greenblatt says that's what keeps the arms race going

Ryan Greenblatt, chief scientist at Redwood Research, has identified a structural problem driving the AI safety crisis: every major lab believes it operates more responsibly than its competitors, creating a race dynamic that nobody can escape.

Greenblatt estimates a 50 to 60 percent risk of AI takeover if development continues on its current trajectory. Philosopher and neuroscientist Sam Harris challenged this assessment by pointing out the comparison fails basic logic. Manhattan Project scientists would have abandoned the atomic bomb effort at a 10 percent existential risk threshold. The fact that the AI industry continues at 50 to 60 percent risk reveals something broken about decision-making at the highest levels.

The core problem sits in how AI companies evaluate their own practices. Anthropic positions itself as the safety-first alternative to OpenAI. OpenAI argues its approach balances safety with capability development. Google, Meta, and other players offer their own safety narratives. From each company's internal perspective, that story holds water. From the outside, the pattern looks different: every lab races ahead while telling itself the others are reckless.

This dynamic creates an impossible situation. If Anthropic unilaterally slows down, OpenAI gains competitive advantage. If OpenAI prioritizes safety, Anthropic or a Chinese competitor pulls ahead. The individual incentive structure pushes toward speed regardless of what leadership believes about responsibility. Greenblatt calls this out directly. The companies cannot self-correct fast enough because the economic and competitive pressures overwhelm safety constraints.

Greenblatt points to international agreements as the only viable escape hatch. A binding treaty structure, similar to nuclear non-proliferation frameworks, could reset the game theory. If all major AI nations agreed to development benchmarks, safety checkpoints, and verification mechanisms, individual labs would face legal consequences rather than competitive disadvantage for moving cautiously. Without that external constraint, good intentions at OpenAI, Anthropic, and elsewhere remain subordinate to race logic.

The timing matters. AI capabilities have reached a threshold where the stakes justify dramatic intervention. Large language models now demonstrate reasoning, planning, and learning abilities that rival or exceed human performance in narrow domains. The gap between current capabilities and autonomous systems capable of self-improvement narrows annually. Greenblatt's 50 to 60 percent risk estimate reflects this acceleration.

The industry response so far has taken two forms. Internally, companies invest in safety research and implement testing protocols. Externally, they lobby against regulation while simultaneously warning governments about AI risks. This two-faced approach signals that current safety measures remain insufficient even from their architects' perspectives.

What changes next depends on whether policymakers internalize Greenblatt's point. If governments continue treating AI safety as an industry self-regulation problem, the race continues. If they recognize the game-theoretic trap and establish binding international frameworks, labs gain legal cover to slow down without losing competitive position. Neither path guarantees safety, but one path at least removes the structural incentive to cut corners on existential risk.