The Machine Intelligence Research Institute has published an informal summary of a paper by its Technical Governance Team titled “AI Governance to Avoid Extinction: The Strategic Landscape and Actionable Research Questions,” originally published in May 2025. The summary describes four geopolitical trajectories the authors believe humanity may follow as AI capabilities increase, and argues that only one avoids catastrophic outcomes including human extinction.
The four trajectories described are: Off Switch and Halt, in which nations cooperate to build technical and legal infrastructure capable of shutting down unsafe AI systems and potentially halting frontier development entirely; US National Project, in which the US races to build superintelligent systems and asserts control over global AI development; Light-Touch, in which governments largely leave AI companies to regulate themselves; and Threat of Sabotage, in which nations respond to rival AI development through sabotage or the credible threat of it.
The post evaluates each trajectory against four categories of catastrophic risk: loss of control, in which AI pursues misaligned goals; misuse, in which malicious or reckless actors weaponise AI; war, in which great-power conflict is triggered by AI competition; and authoritarianism or societal lock-in.
MIRI’s summary states the Off Switch and Halt trajectory is “the best path to mitigate loss of control risks from AI,” and devotes the most attention to it. The authors describe an “off switch” as the technical, legal, and institutional infrastructure required to shut down AI systems and verify that no dangerous AI development is occurring globally. They argue this could be built without requiring global consensus on whether a halt is ultimately necessary, and that doing so would also reduce risks from terrorism and geopolitical destabilisation.
The summary describes several potential verification and enforcement mechanisms for a halt: physical inspections of datacentres, chip-level location verification via network latency measurement, restrictions on the publication of algorithmic improvements, and redirecting key researchers toward safer activities. The authors note that current chip security measures are insufficient to stop a well-resourced attacker from spoofing location data, and that secure chip infrastructure remains an active research target.
On the US National Project trajectory, the post describes potential benefits — US control might prevent authoritarian lock-in by a rival power — but notes that it would still leave open the risk of misaligned AI and could accelerate a global arms race. The Light-Touch trajectory is described as likely producing competitive dynamics where safety is systematically deprioritised. The Threat of Sabotage trajectory is described as potentially destabilising and prone to miscalculation.
The post links to the paper’s executive summary and full report for readers seeking a more technical treatment.