Prohibition vs. Traceability: The Divergence of Two AI Governance Approaches in China and the U.S.

—Two Governance Paradigms: The U.S. Ban on Artificial Superintelligence Act and the Wuzhen Summit

In November 2026, two nearly simultaneous developments brought the fork in the road for artificial intelligence governance into plain view.

On one side is the U.S. Congress. After the midterm elections, the Democratic Party took control of both the Senate and the House of Representatives, and the Ban on Artificial Superintelligence Act jointly submitted by Sanders and Khanna entered the legislative spotlight—advanced models trained with computing power reaching 10 operations would be subject to a moratorium on research and development, individuals in violation could face up to 20 years in prison, companies could be forcibly dissolved, a new cabinet-level Federal Department of Artificial Intelligence would be established, and research and development entities would be required to report within 24 hours once they detect "signs of superintelligence," with the entire set of standards packed into the toolbox of foreign policy and export controls.

On the other side is the Wuzhen Summit of the 2026 World Internet Conference in Wuzhen, Zhejiang. 138 countries and regions, 1,800 guests, and 670 enterprises gathered in the water town to discuss open-source collaboration, AI trustworthiness, technology for good, and data security—and how to find that dynamic balance point between the speed of innovation and the bottom line of security.

Same year, same issue, two nearly opposite lines of thinking. This is not simply a question of "who is stricter and who is looser," but a head-on encounter between two sets of underlying assumptions about where risk comes from, where the gate should be placed, and who should hold the gate.

I. The Prohibition Approach: Anchoring Risk at the Entry Point

The internal logic of this U.S. bill is actually quite clear: it assumes that frontier risk is a single-point explosion—as long as computing power crosses a certain line, superintelligence may suddenly be born in a certain training run, and once born, it is irreversible. Since danger arises only at the moment of ignition, the most effective way is to take away the matches before ignition.

Thus governance is compressed into a single gate: draw the line, pause, impose heavy penalties, require reporting, and control spillovers. Administrative power stands at the entry point of technology, trading "not allowed to do it" for security.

The advantage of this design is that it is unambiguous. The rule is singular, the indicator is quantifiable, and the deterrent effect is extremely strong. Regulators do not need to understand what is happening inside the model; they only need to check a number. For legislators, this is the easiest plan to explain to voters; for law enforcers, this is the lowest-cost lever.

But it stakes everything on one premise: that risk occurs at only one point, and that point can be accurately marked by a number.

II. The Traceability Approach: Building Security into the Process

China's path goes in another direction. It does not presuppose that beyond a certain line nothing may proceed, but requires that no matter how far one goes, every step must leave a trace.

Where data comes from, what processing it undergoes, how the model evolves, how the reasoning chain proceeds, what the basis for output is, how anomalous signals are identified, and whether humans can take over at critical moments—these links constitute a verifiable chain. Combined with tiered and classified management, sandbox pilots, open-source collaboration, and equal emphasis on development and security, governance does not stand at the door turning people away, but is embedded within the system and operates continuously.

The underlying assumption of this design is completely different: risk is not born only at the moment of training, but continuously emerges during fine-tuning, alignment, autonomous scheduling by agents, multi-model interaction, and long-term continuous operation. Since risk is distributed and runs through the entire life cycle, governance cannot merely set up a gate at the entry point; it must make the entire process visible, measurable, and intervenable.

Summarize the difference between the two in one sentence.

The prohibition approach asks: How large an AI is not allowed to be built?

The traceability approach asks: No matter how large it is built, can we understand it, control it, and prove that it is safe?

III. The Essence of the Divergence Lies at Three Levels

The first level is a different imagination of risk. One side regards frontier models as a single entity that may suddenly spiral out of control, while the other regards them as a complex system that continuously evolves and constantly generates new behaviors. The former needs a red line; the latter needs a net.

The second level is a different governance lever. Prohibition relies on administrative power to impose constraints from outside, and the effectiveness of regulation depends on whether the regulated truthfully report; traceability relies on the technical system to prove its own innocence, and the effectiveness of regulation depends on whether the chain is truly verifiable. The former is "power regulating technology"; the latter is "technology carrying inherent regulability."

The third level is a different power structure. The prohibition approach naturally tends toward centralization: determination standards are held by the government, exemptions are allocated by the government, and international rules are led by a few countries and spill outward. The traceability approach naturally tends toward decentralization: standards can be co-built by multiple parties, capabilities can be adapted by tier, and economies at different stages of development can all find their own place.

IV. Three Inherent Paradoxes of the Prohibition Approach

This seemingly toughest plan is precisely vulnerable in three places.

First, thresholds can be circumvented by technology. The computing power red line is a static indicator, while technological progress is dynamic. Splitting one ultra-large training run into multiple rounds of incremental pre-training, using distillation compression, splitting with mixture of experts, using distributed asynchronous collaboration, and using small model clusters for relay evolution—once the statistical basis for computing power in a single training run is decomposed, that line of 10 loses its binding force. Rules run behind technology; the harder the rules, the faster circumvention schemes iterate.

Second, heavy penalties suppress reporting, which instead blinds regulation. If "identifying signs of superintelligence" itself lacks objective and verifiable determination standards and relies heavily on human subjective judgment, then research and development entities face a highly asymmetric choice: reporting means the project immediately stalls, the institution faces punishment, and the responsible person bears criminal liability; not reporting means hoping not to be discovered. Under uncertainty, the rational choice is often not to proactively expose anomalous signals, but to dissolve them within the system. This mechanism, intended for early warning, may in turn weaken early warning capability.

Third, rules are easily geopoliticalized. Writing computing power standards into diplomatic objectives and export controls means that it is not only domestic security legislation, but will also become a bargaining chip in technological competition. Using a self-defined "superintelligence threshold" to screen other countries' chips, clusters, and model cooperation will wrap governance issues into the logic of competition, ultimately damaging the collaborative foundation for the world to jointly address risks.

V. The Practical Thresholds of the Traceability Approach

Of course, the traceability approach is not without cost; it is much harder than prohibition.

It requires a real technological foundation. Traceability is not writing a compliance document, but truly growing into the system a whole set of capabilities including data sources, training logs, weight changes, reasoning paths, risk alerts, and manual circuit breakers. Without this foundation, "traceability" remains only a slogan.

It requires unified standards and third-party verification. Log formats that each party defines on its own cannot be mutually recognized, nor can regulators compare them. In the absence of standards, traceability degenerates into each party presenting evidence and each saying its own thing.

It requires controllable costs. If only a few giants can afford a complete audit and monitoring system, then compliance itself will become a new market barrier, small and medium-sized institutions will be squeezed out, and diversity of innovation will be lost accordingly.

It also needs to prevent formalism. If the ledgers are beautifully made and the chain records are complete, yet no one truly looks at them, verifies them, or intervenes based on them, then even the most complete traceability is merely decoration.

VI. Conclusion

The two approaches appear to be arguing about "strict or not," but in fact they are arguing about a more fundamental question: should security be written into legal provisions, or grow into the system.

The prohibition approach trades certainty for security—rules are simple and deterrence is clear, at the cost of compressed innovation space, endless technological circumvention, and governance easily hijacked by geopolitics. The traceability approach trades visibility for security—allowing movement forward, but requiring every step to leave a trace, be auditable, and be stoppable, at the cost of high technological thresholds, slow standard-building, and heavy implementation costs.

Facing a technology whose capabilities far exceed expectations, humanity instinctively wants to find a gate. But history has repeatedly proven that truly effective security has never been about locking the door shut, but about making everything in the room visible, every anomaly reportable, and every switch in human hands.

Global AI governance will ultimately not have only one answer; countries differ in development stages, legal traditions, and industrial landscapes. But one point will probably be verified by time: whoever can build security capabilities into the technological foundation rather than merely writing them into legal provisions is more likely to both run forward and stop in time.

Tuogen Institute of Artificial Intelligence

Domestic Trustworthy and Controllable AI Foundation Laboratory

Chen Jiqiang

November 8, 2026