What would it actually take to move from an AI arms race to “One AI”?
The idea of a globally governed frontier AI system raises a much harder question than how such a system might be built:
How would we govern it?
My latest article examines three challenges that any serious “One AI” proposal has to confront:
- Enforcement: How can international cooperation be made credible when the incentives to defect may be enormous?
- The path from current AI security failures to future alignment risk: How do we establish the causal connection without simply jumping from today’s sandbox failures to speculative catastrophe?
- The open-weights paradox: How can we preserve openness, independent scrutiny and broad access while preventing highly capable models from becoming easily weaponised?
The argument is that “One AI” is ultimately less a technology proposal than a governance problem.
If the stakes of advanced AI really are as high as many increasingly argue, then governance cannot be an afterthought. It has to be designed alongside the technology.
Read the full article:
From Superintelligence Arms Races to One AI
Can frontier AI be governed globally without sacrificing either openness or security?
First dropped: | Last modified: September 15, 2026