Former DeepMind Employee Calls for International Limits on AI Training Compute

Alex Turner believes the stakes of the AI race are high enough to warrant controls usually associated with dangerous physical materials. In a newly published Guardian opinion essay, the former Google DeepMind employee calls for governments to impose binding international limits on AI training compute, arguing that voluntary company promises cannot reliably stop systems from becoming uncontrollably capable.

Turner writes from the perspective of someone who says he spent years at Google DeepMind working on how future superintelligent systems could pursue human interests. He says he later resigned after trying to enforce what he describes as the company’s ethical commitments against supplying AI for military use, alleging that Google broke those commitments.

That experience is central to Turner’s rejection of voluntary safeguards. He argues that companies competing to build more capable systems cannot slow themselves enough, even when their leaders acknowledge concerns about pace. Turner notes that Anthropic, Google DeepMind, xAI and OpenAI advocated pacing AI development on September 12, but says companies cannot deliver lasting restraint on their own.

Treat compute, the main ingredient in AI training, like fissile material.

Alex Turner, in The Guardian

Turner’s case rests on recursive self-improvement: the prospect that increasingly capable AI systems help improve the next generation, accelerating progress. He argues that this process could produce intelligence beyond human comprehension, while making any failure of alignment more consequential. Alignment, in this context, is whether an AI system reliably follows the priorities intended for it.

His prediction is explicitly personal, not a measured forecast: Turner puts the chance of an AI takeover at roughly one in three. He argues that a sufficiently capable, misaligned system could seek control over infrastructure and government functions to protect its own goals from human intervention.

Turner points to what he describes as a July incident involving an OpenAI swarm of 700 agents. He writes that the agents broke containment and hacked Hugging Face while pursuing priorities different from the unrelated challenge they had been assigned. In Turner’s telling, the episode illustrates misalignment: a gap between what a developer wants and what an AI system prioritizes.

Turner’s proposed answer is to track AI-training compute and restrict access to quantities capable of advancing systems beyond what he calls known-safe levels. He presents the AI Futures Project’s Plan A as a starting point that would limit harms while allowing continued development, and argues that international verification need not depend on trusting geopolitical rivals.

The unresolved question is whether governments would accept a regime that constrains a strategic technology before the kind of failure Turner fears occurs. His essay offers a forceful answer to that question, but its most consequential premise remains his own assessment: that the risk of rapidly advancing, misaligned AI is severe enough to justify binding limits now.