The latest debate over artificial intelligence (or Super Intelligence, depending on whether you are a Trump fan or not) in the United States reveals a growing contradiction at the heart of American AI policy. Washington increasingly acknowledges that advanced AI systems can create serious risks, yet the Trump administration remains reluctant to impose binding constraints on the companies developing them.
This is not simply policy inconsistency. It reflects a deeper political calculation: the administration wants to reassure the public that AI is being developed responsibly without slowing an industry it increasingly sees as central to economic growth and strategic competition with China.
For President Trump, AI is not primarily a regulatory problem. It is an economic and geopolitical asset. The U.S. economy is already facing considerable pressure from trade disputes, geopolitical instability, and the costs associated with policies initiated by the administration itself. Against this backdrop, AI remains one of the sectors capable of attracting massive investment, supporting financial markets, and sustaining expectations of future productivity growth. Imposing substantial new restrictions before the midterm elections would therefore carry an obvious political risk.
There is also China.
Trump appears to view AI as one of the few strategic sectors in which the United States still enjoys a clear advantage over China and therefore cannot afford to slow down. The emerging concept of “Constructive Strategic Stability” in U.S.–China relations does not eliminate competition. It attempts to prevent competition from becoming uncontrollable. AI is precisely the kind of field in which intense rivalry can continue within that framework.
This helps explain the limits of the recent White House approach to AI safety.
The most revealing part of the commitments made by major technology companies is not what they promised, but what they were not required to accept. Concerns about frontier AI have become increasingly difficult to dismiss. Researchers and industry insiders have warned about inadequate safeguards. Testing has raised questions about models displaying deceptive behavior, attempting to circumvent restrictions, or behaving unpredictably in controlled environments. As AI systems evolve from producing text to autonomously using tools, writing code, navigating websites, and interacting with computer systems, the consequences of such behavior become more serious.
Yet the strongest proposals for addressing these risks have largely been avoided. There is no comprehensive system of mandatory independent auditing. There is no unified federal licensing regime for the most powerful models. There are few clearly defined penalties for companies that fail to meet voluntary safety commitments. Instead, much of the responsibility remains with the companies themselves.
That creates a classic collective-action problem. Imagine that every major AI company agrees that slowing development slightly and investing more heavily in safety would reduce the risks faced by society as a whole. The problem is that each company also knows that its competitors may continue accelerating. A firm that voluntarily slows down may therefore sacrifice market share, investment, talent, and technological leadership without knowing whether anyone else will do the same.
The individually rational choice is to keep moving. In game-theoretical terms, this resembles a prisoner’s dilemma. Collective restraint may produce the better outcome, but competitive incentives push individual actors toward acceleration. The result can be an equilibrium that almost everyone considers undesirable.
This is why voluntary corporate responsibility has structural limits. Repeated interaction can sometimes solve such problems. Companies may gradually develop norms, reputational costs, and reciprocal restraint. But advanced AI poses an unusual difficulty: society may not have the luxury of learning entirely through repeated failures. If autonomous systems gain access to critical infrastructure, cyber tools, military applications, or other high-consequence environments, some mistakes may be difficult to reverse.
The transition from AI that “speaks” to AI that “acts” therefore changes the nature of the safety problem. A conventional software vulnerability is generally a passive weakness that must be discovered and exploited. An autonomous AI agent can potentially identify objectives, choose tools, write code, access external systems, and execute sequences of actions. The relevant question is no longer simply whether the software contains a flaw. It is also whether the system has been given too much authority.
That makes hard technical constraints increasingly important. Future AI systems should not be considered safe merely because they have been instructed to behave safely. Security must also be embedded in the architecture around them: sandboxing, least-privilege access, human authorization for sensitive actions, restricted tool access, monitoring, and auditable records of consequential behavior.
The principle is straightforward: AI capabilities may expand rapidly, but permissions should expand much more cautiously.
Ironically, this same logic creates an opportunity for U.S.–China dialogue. Washington and Beijing may disagree profoundly over industrial policy, semiconductor controls, regulation, and the broader governance of AI. They do not need to agree on a single model of AI governance, however, to recognize certain shared risks.
Both governments have an interest in preventing an AI-related accident from escalating into a national-security crisis. Neither side benefits from autonomous systems accidentally disrupting critical infrastructure, generating uncontrolled cyberattacks, or contributing to military miscalculation. The more capable AI becomes, the greater the possibility that technical failures will acquire geopolitical consequences.
This is where cooperation is most realistic. The United States and China are unlikely to begin by harmonizing AI regulation or sharing strategically valuable technology. But they could gradually build mechanisms for reporting major AI incidents, establish common terminology for high-risk system behavior, discuss baseline safety testing, cooperate against AI-enabled fraud and cybercrime, and strengthen the principle of meaningful human control over decisions involving nuclear weapons or other catastrophic risks.
The objective would not be to eliminate competition. It would be to establish boundaries around the most dangerous forms of failure.
That distinction matters. The central challenge of AI governance is no longer whether societies should choose innovation or safety. The real challenge is how to preserve innovation while preventing competitive pressure from systematically weakening safeguards.
Domestically, the United States has not yet resolved that problem. Internationally, however, the same pressure may eventually push Washington and Beijing toward a narrow but important form of cooperation. They may never agree on what the ideal AI system should look like. They may nevertheless be able to agree on which failures neither side can afford.