Ethereum co-founder Vitalik Buterin said public debate on AI often fixates on whether machine intelligence itself will become dangerous, while paying less attention to a separate issue with wider political consequences: what happens if control over advanced AI systems ends up in the hands of a small number of actors.
He argued that if AI eventually outperforms humans at all critical tasks, humanity’s collective bargaining power could fall to zero. That would not be a narrow technical shift. It would alter the global balance of power. In his view, scenarios where a few AI labs dominate development and deployment deserve serious concern on their own; outside the technology sector, such concentrated power would already be treated as unacceptable.
The debate shifts from intelligence to control
Buterin warned against assuming that the transition to advanced AI will be smooth and fairly managed. He described that belief as risky in itself, and not examined closely enough. He also questioned why proposals for open debate, and even for acceleration under public scrutiny, are often met with more suspicion than the assumption that centralized governance can safely manage future AI risks.
The point of his argument is structural. The key question is not only how capable AI becomes, but who controls training, deployment, and the ability to stop large-scale systems when conditions change.
Open-source access and shared emergency brakes
As a practical response, Buterin pointed to updated provisions in major AI roadmaps that require open-source access in AI development. He described those mandates as an important move against monopolization. Broader access to foundational models and code is meant to reduce the chance that a handful of companies capture technologies with system-wide impact.
He also referenced the idea of mutually assured compute destruction, a governance mechanism under which multiple parties would agree in advance to pause or slow large-scale AI computation under specified conditions. The design aims to distribute intervention power rather than concentrate it, creating checks that prevent any one actor from shutting others out.
For Buterin, that kind of arrangement is preferable to frameworks where select groups can exclude opponents or dissenting voices from major governance decisions.
d/acc as a decentralized defensive model
Buterin tied that view to his support for d/acc, a decentralized defense strategy aimed at AI-related existential threats. The priorities he listed include formal verification, advanced cryptography, secure open hardware, pandemic readiness, and food security.
He also included stronger public epistemics, meaning systems that help societies maintain reliable shared facts, along with preparation for risks that could appear whether or not AI advances at the pace some expect. That places the discussion well beyond software capability alone and into the resilience of social coordination.
Pre-agreed triggers for disputes over AI speed
Buterin suggested that disagreements between advocates of rapid AI progress and those calling for caution could be mediated through pre-agreed triggers. He gave examples such as the onset of severe pandemics or unemployment rising above 25%. If those conditions occur, or fail to occur within an expected window, legitimacy could shift toward whichever side made the more accurate prediction.
He said participation in these decisions should not be limited to governments and corporate executives. Wider community involvement, including discussion on public platforms and social media, should be part of the process. While acknowledging that every proposal can be criticized as unrealistic, he said he welcomes continued efforts to design governance systems, especially ones aimed at reducing concentrated control over AI development.