There Is No Arrow: AI Strategy, US-China Competition, and the Limits of Linear Thinking
There Is No Arrow
A recent conversation on the Just Security podcast is worth sitting with. Three serious analysts -- Brianna Rosen, Jenny Marron, and Sam Winter-Levy -- were discussing the US AI Action Plan, and they kept doing something that is both refreshing and unsettling. They refused to resolve the tensions they were naming.
On cybersecurity, the problem was simultaneously technical and organizational, and separating them produced bad policy. On export controls, the administration was simultaneously committed to AI leadership and undermining it, and both things were true at the same time. On market dominance versus compute advantage, US companies operating overseas were simultaneously an asset and a vulnerability, and the right answer depended entirely on which theory of competition you found more convincing.
No single causal arrow pointed from here to there. Everything was conditioning everything else.
Mutual Causality
A recent post on this Substack wrote about what is happening in Minneapolis right now. Fifty thousand people marching in minus 20 degrees. Two hundred thousand diapers moving through informal networks. Grocery stores delivering free to families who will not leave home. No funder designed this. No nonprofit manages it. No logic model predicted it.
The concept that emerged to explain it was Joanna Macy’s mutual causality. Macy, a Buddhist systems theorist, observed that two traditions separated by millennia had arrived at the same insight: things do not cause each other in a straight line. They arise together. They condition each other. When conditions are sufficient, the thing emerges, not because one input produced one output, but because everything was ready simultaneously.
The logic model assumes linearity. Input, activity, output, outcome, impact. Arrow, arrow, arrow. What happened in Minneapolis is not linear. Neither is what the analysts were describing on that podcast. US AI strategy is not a logic model. It is a web of conditions: trade pressures, compute advantages, alliance relationships, domestic politics, and the behavior of a rival whose intentions cannot be fully read, all conditioning each other at once. There is no single arrow that points from current policy to a stable outcome.
The Window of Vulnerability Problem
The central unresolved question in US-China AI competition is this: under what conditions does capability asymmetry increase the risk of conflict rather than deter it?
The nuclear analogy that dominates strategic thinking assumes capability gaps produce stable deterrence. The stronger party wins without fighting because the weaker party knows it cannot win. But AI may not behave that way. A country falling behind faces incentives to act before the gap widens, particularly if it believes the leading country will eventually convert that advantage into coercion. This is the window of vulnerability problem, and the field does not yet have good models for how AI capability trajectories interact with crisis stability.
It matters enormously because the policy interventions that follow from different answers are almost opposite. If asymmetry produces deterrence, the right strategy is to maximize and maintain the US lead at all costs. If asymmetry produces instability, the right strategy includes some degree of capability transparency or cooperative constraint even with an adversary. The analysts on the podcast were circling this question without quite naming it. Their careful hedging was not analytical weakness. It was an honest acknowledgment that deterrence and instability may not be competing dynamics at all. They may be the same web, seen from different angles.
If that is true, then the search for the right answer to whether asymmetry deters or destabilizes is itself the wrong question. The right question is what conditions make the web more or less likely to produce catastrophic outcomes, regardless of which dynamic is operating.
Two Interventions in Tension
When thinking about feasible policy responses given this uncertainty, two interventions emerge that pull in opposite directions, and pursuing both simultaneously may be the only honest response. Neither of these ideas is new. What is less common is being honest about why they are in tension with each other and pursuing both anyway.
The first is a bilateral AI incident reporting mechanism between the US and China. Not a comprehensive agreement. Not a treaty. A narrow, technical channel for reporting AI related incidents that could be misinterpreted as hostile acts: system failures, unexpected autonomous behaviors, cyberattacks involving AI components. The diplomatic infrastructure for this exists. The 2023 San Francisco summit reopened channels that could support exactly this kind of mechanism. The precedent is the nuclear hotline after the Cuban Missile Crisis: not an expression of trust, but a practical tool to prevent miscalculation when trust is limited.
The second is a G7 commitment to publish standardized, unclassified capability assessments of frontier AI models on a regular cadence. Western decision makers currently lack reliable, comparable information about the trajectory of Chinese frontier model development. That gap produces both overreaction and underreaction, sometimes simultaneously in different parts of the same government. Building on existing NIST and UK AISI evaluation frameworks, this is achievable within a year through executive coordination among G7 governments without requiring Chinese participation or a new multilateral institution.
Both interventions depend on diplomatic infrastructure that has been significantly disrupted over the past year. The bilateral channel assumes US-China relations stable enough to sustain a narrow technical agreement, at a moment when the trade war, export control reversals, and broader geopolitical friction have strained those relations considerably. The G7 mechanism assumes allied coordination at a moment when the current administration has actively unsettled those relationships. These are not arguments against pursuing either intervention. They are reminders that policy proposals exist inside conditions, not above them. The web includes the diplomats and the institutions and the political will, and right now several of those threads are frayed.
The tension between the interventions themselves is also real and worth naming directly. The incident reporting mechanism assumes enough shared interest in stability to sustain a cooperative technical channel. The capability assessments risk undermining that assumption. China is likely to read standardized Western assessments of Chinese models as competitive surveillance rather than neutral information sharing. The first intervention operates from a managed competition frame: adversaries can build narrow cooperative mechanisms to reduce catastrophic risk. The second operates from a strategic competition frame: Western governments need better intelligence about China to maintain advantage.
Both are worth pursuing anyway, for the same reason arms control negotiations continued alongside intelligence collection during the Cold War. The instability risk from advanced AI is severe enough that incident reporting mechanisms are worth preserving even under competitive conditions. But a diplomat pursuing the first intervention would need to carefully manage how the second is framed and communicated to Chinese counterparts. The tension does not disappear. It has to be actively managed.
There Is No Arrow
This is mutual causality applied to geopolitics. The conditions shaping US-China AI competition are not arranged in a neat causal sequence. They are a web: capability trajectories that cannot be fully measured, intentions that cannot be fully read, domestic pressures on both sides that constrain what leaders can agree to even when agreement might serve their interests, and an underlying question about whether asymmetry stabilizes or destabilizes that the field cannot yet answer.
What the analysts on that podcast were modeling, whether they named it this way or not, is what it looks like to make policy inside a complex system. Contradictory interventions get held simultaneously. Hedging is not a failure of conviction. It is a response to genuine uncertainty. The pressure to draw a single arrow from cause to effect gets resisted when no such arrow exists.
In the near term, the most honest strategy is to keep the channels open that still exist, invest in the evaluation infrastructure that does not require Chinese cooperation, and resist the pressure to declare a winner between managed competition and strategic competition before the evidence supports it.
In Minneapolis, the conditions for what is happening now were laid down over decades, activated by crisis, maintained through quiet, and reactivated when the moment demanded it. No one predicted it. No logic model contained it.
AI strategy is the same. The conditions are accumulating. The web is tightening. There is no arrow from here to a stable outcome. There are only conditions that can be shaped, tensions that can be managed, and the discipline to resist false clarity when the situation does not offer it.
That is not a counsel of despair. It is the beginning of honest strategy.
Anthralytic is a strategy and evaluation studio helping mission-driven teams clarify and amplify their impact through human expertise, data, and AI. We work at the intersection of social change, emerging technology, and accountability.
