M2-A1
General AI capabilities enter multiple domains → capabilities with major consequences spread
AI enters knowledge work, agents, military activity, cyberspace, infrastructure, decision support, and other settings simultaneously. M2's first change is not simply that AI becomes smarter, but that the same class of general capabilities begins to be reused across multiple systems with major consequences.
Real-world structural assessment
Boundary: Capabilities entering multiple domains do not directly imply that civilizational risk must rise. Permissions, organizational adoption, task boundaries, and consequences of failure require further assessment.
M2-A2
Capabilities are reused across domains → state/organizational competition absorbs AI
When AI is seen as a common underlying capability for advantages in economics, intelligence, cyberspace, unmanned systems, and command, states and organizations incorporate AI adoption into considerations of relative advantage and security. Even without knowing whether more AI necessarily improves safety, they may fear that non-adoption will first create a disadvantage.
Strategic-mechanism assessment
Boundary: Competitive absorption is not technological destiny. Opponents may accelerate, but they may also supervise, decline use, verify independently, or establish shared guardrails.
M2-A3
Competition absorbs AI → pressure for speed, autonomy, and prior authorization rises
In competition with major consequences, faster perception, proposal generation, and execution create practical advantages. If communications are vulnerable, waiting is costly, or organizations fear losing the ability to act, pressure arises to shorten verification, expand prior authorization, and increase autonomy when disconnected.
Conditional causality
Boundary: Speed is not itself the risk. The key question is whether verification, communication, renewed political decision-making, and stopping capabilities remain available alongside it.
M2-A4
Greater speed/autonomy/prior authorization → windows for immediate human confirmation and correction may shrink
When machine action outpaces human review, permissions expand into real-world execution, and irreversible actions are difficult to recall, a formal “human in the loop” may no longer mean that humans retain actual control.
Assessment of risks with major consequences
Boundary: Distinguish currently deployed systems, experimental systems, and prospective scenarios; do not describe all AI systems as already acting automatically beyond human control.
M2-A5
AI enters perception/judgment/action chains → attribution, deception, and uncertainty change structurally
More data and stronger models may improve situational awareness, but may also create new uncertainties through data contamination, adversarial manipulation, synthetic content, model errors, and opaque provenance. More information does not automatically reduce misjudgment.
Strategic-epistemological assessment
Boundary: Difficult attribution is not unique to AI. AI may also improve anomaly detection, cross-checking, and crisis communication; the net effect requires comparison.
M2-A6
Software diffusion, civilian–military dual use, and infrastructure dependence → capabilities spread faster, but not equally
AI software, models, and civilian infrastructure can spread across organizations, lowering barriers to some tasks. But actual capabilities with major consequences remain constrained by computing resources, data, talent, hardware, organizational absorption, and task understanding. Faster diffusion therefore does not mean equal capability among all actors.
Diffusion-mechanism assessment
Boundary: Model availability must not be equated directly with strategic-level capability; unequal capabilities must not obscure low-cost tasks with high impact.
M2-A7
Competition, autonomy, speed, attribution, and diffusion change together → local security logics may be repeatedly executed/amplified by stronger intelligence
The central concern is not AI suddenly becoming malicious. It is that logics of efficiency, goal achievement, security, self-protection, and friend–enemy classification, which have reasons within local human contexts, may be executed over long periods by more capable, faster systems with broader permissions, and potentially abstracted into more general strategic relationships.
Civilizational-interface assessment
Boundary: This is a conditional structural inference, not a factual claim that current AI has developed a unified value system, friend–enemy worldview, or will as a subject.
M2-A8
Local logics acquire stronger intelligent capabilities → the old civilization's margin for correction begins to change
M2 establishes only that historical conditions are changing: capability, speed, permissions, diffusion, and systemic dependence begin to alter the reparability of mistakes within the old civilization. Whether this makes gradual upgrading insufficient requires M3 to test independently whether cross-domain coupling, irreversibility, and governance feedback actually exceed the old safety margin.
Bridge between nodes M2 → M3
Boundary: M2 cannot directly establish that Civilization Leap is necessarily the only way. M3 must independently show when old governance becomes mismatched and which counterexamples would overturn that assessment.