AI Safety Researcher Puts Takeover Risk at 50 to 60 Percent, Blames Lab Race Dynamics
Ryan Greenblatt of Redwood Research warns that competition between AI labs like Anthropic and OpenAI is driving an arms race, estimating a 50 to 60 percent chance of an AI takeover if development continues on its current path.
Ryan Greenblatt, chief scientist at the AI safety organization Redwood Research, estimates the probability of an AI takeover at 50 to 60 percent if development continues on its current trajectory. In a conversation with Sam Harris, Greenblatt argued that the very labs most concerned about safety are locked into a competitive dynamic that makes it difficult for any single actor to slow down.
Greenblatt's central claim is that every major AI laboratory believes it is the responsible one. Each lab, he suggests, sees itself as the best steward of powerful technology and therefore concludes that it must stay at the frontier to prevent less careful competitors from setting the pace. That shared conviction, rather than any single reckless actor, is what keeps the arms race going. He pointed specifically to the race dynamics at Anthropic and OpenAI, two of the most prominent developers of advanced AI systems, as examples of how safety-focused intentions can coexist with accelerating competition.
Sam Harris pushed back on the figure, noting that it does not square with the speed at which the industry is moving. Harris observed that scientists working on the Manhattan Project would likely have halted their work if they had believed there was even a 10 percent chance of catastrophic consequences. The comparison highlights a striking gap between the scale of the risk Greenblatt describes and the continued rapid pace of commercial AI development.
Greenblatt's answer to this dilemma is not to rely on voluntary restraint by individual companies. Instead, he is counting on an international agreement to impose binding limits on the most dangerous forms of AI development. Such a treaty, in his view, would change the incentives that currently push labs to compete rather than coordinate. Without it, he sees little reason to expect that any single lab will unilaterally step back, because doing so would simply cede the frontier to others.
The discussion touches on a broader debate within the technology and policy communities about how to govern artificial intelligence. Critics of the current approach argue that safety rhetoric from labs often serves to legitimize continued scaling, while proponents of international coordination insist that only a shared framework can prevent a race to the bottom. Greenblatt's position sits between these camps: he is deeply skeptical of the industry's ability to self-regulate, yet he remains hopeful that a diplomatic solution can be reached before the risks materialize.
For now, the numbers Greenblatt cites are a stark reminder of how high the stakes have become. A 50 to 60 percent chance of an AI takeover is not a prediction of inevitability, but it is far above the threshold that would have stopped earlier generations of scientists. Whether the international community can act before that threshold is tested remains an open question.
2
