This is the first part of an analysis of China and global cyberspace governance.
IN RECENT years, global cyberspace governance has remained broadly stable but has shown signs of localised disorder.
Deepfake technologies blur the boundary between truth and fabrication; autonomous agents operating across jurisdictions often generate unintended consequences; malicious activities automatically generated by artificial intelligence (AI) models are increasingly difficult to prevent; and the widespread adoption of generative technologies poses serious challenges to the information ecosystem and social trust. These phenomena all exhibit cross-domain, non-subjective, and non-attributable characteristics.
Action chains naturally transcend geographic boundaries, while key stages often lack clearly identifiable human decision-makers. Responsibility therefore becomes fragmented and drifting among models, computing power, data, and platforms, leaving existing governance systems structurally ill-equipped to respond effectively.
The root cause of these challenges does not lie in the inherent danger of any particular technology. Rather, it lies in the fact that AI is fundamentally reshaping the ontology of cyberspace.
The ontological transformation
Since the birth of the Internet, global cyberspace has long been understood as an “information interaction space” centered on human actors. Its governance system has been built upon several core assumptions, namely identifiable subjects, attributable actions, and predictable rules.
However, with the widespread adoption of large models, intelligent agents, and autonomous systems, cyberspace is gradually transforming from a passive environment for information exchange into a “global intelligent action domain” jointly constituted by humans and AI, driven by algorithmic systems, and characterised by continuously generated actions.
The essence of today’s governance dilemma lies precisely in the ontological rupture between governance paradigms and the actual operational logic of cyberspace.
With AI deeply embedded in cyberspace, the latter is shifting from an “information interaction space” centered on human communication to a “global intelligent action domain” in which humans and machines jointly participate. The “global intelligent action domain” refers to a dynamic action generation structure composed of humans, AI systems, autonomous agents, and underlying infrastructures.
Within this structure, actions exhibit generative, diffusive, and evolutionary characteristics. The system remains in a state of continuous flux, making it difficult to statically record or retrospectively predict. Power relations are deeply embedded in model architectures, interface standards, and ecological mechanisms. Cyberspace thus evolves from a passive channel for information transmission into an environment in which actions are generated and propagated.
Breaking it down
This ontological transformation of cyberspace in the age of AI can be analysed from existential, temporal and power dimensions, corresponding respectively to the sources of action, the evolution of action chains and the distribution of power within the action environment -
Existential reconstruction: As AI models and autonomous agents become embedded in the network layer, platform layer, and application layer, they increasingly execute decisions, generate content and shape interaction patterns within cyberspace. They thus emerge as new actors capable of exerting agency. As a result, cyberspace is shifting from a human-centered domain to a complex ecosystem of human-machine coexistence, and its ontological boundaries are being redefined.
The most prominent manifestation of this existential reconstruction is the autonomisation of action generation. Large language models can generate content, codes, and strategies, while autonomous agents continuously adjust their behavior based on goal settings and environmental feedback.
Temporal reconstruction: Traditional governance systems rely on a basic temporal assumption that the states of them can be recorded, frozen and traced retrospectively. Actions are understood as discrete events occurring at identifiable moments in time, and their consequences can be traced and adjudicated ex post. Legal systems therefore largely rely on mechanisms of ex post accountability.
In the intelligent action domain, however, this temporal structure undergoes fundamental changes. Through continuous fine-tuning, online learning, retrieval augmentation, and parameter updates, AI systems remain in a state of constant evolution.
Power reconstruction: In the traditional Internet order, power is primarily manifested in the control of “entry points,” such as cross-border data gateways, platform rule-making authority, or network access management. States, platforms, and other actors competed around access rights, forming a governance structure centered on “gatekeeping power.” In the intelligent action domain, however, this power structure has shifted significantly.
First, the centre of power moves upward from the data entry point to the model layer. Model architectures, training corpora, alignment mechanisms, and system interfaces collectively shape the behavioral logic and action tendencies of AI systems, thereby determining which actions can be generated and recognised by the system. To control the model is to hold the power to define the logic of action.
Secondly, power increasingly manifests as “ecological power” shared by corporations, platforms, and technological communities. Ecological power includes the control of API openness or closure, rules for model interoperability, control over datasets and training processes, agent deployment policies, and transparency mechanisms. Compared with traditional “gatekeeping power,” “ecological power” more deeply determines the rules, boundaries, and evolutionary trajectory of the action domain.
Finally, the power structure of the intelligent action domain displays a distributed configuration. Providers of computing power, platform operators, model developers, open-source technical communities, and state regulatory institutions each control different key nodes. Power thus becomes networked and difficult for any single actor to monopolise. This structure creates a misalignment between legal governors, such as states, and de facto governors, such as platform enterprises and model developers. Consequently, power in the intelligent action domain no longer lies primarily in controlling information entry points but in shaping generative mechanisms and ecosystem structures.
Structural impacts on global cyberspace governance
The ontological transformation of cyberspace brought about by AI directly challenges the global cyberspace governance system, which is fundamentally predicated on the assumptions that actions are locatable, traceable, and predictable.
As cyberspace evolves into an intelligent action domain, these assumptions gradually collapse, generating structural tensions across three dimensions, namely jurisdiction, rule supply and governance effectiveness.
When it comes to existential transformation, action chains and the applicability of jurisdiction, through the context of jurisdictional allocation, global cyberspace governance has long relied on identifiable connecting points such as the place where an act occurs, the place where service is provided, and the place where harm materialises.
Through extraterritorial application of criminal law, international cooperation clauses in cybercrime conventions, and geographic information such as domain names, IP addresses, and server locations, states have developed a division of enforcement responsibilities for transnational cyber activities. During the “information interaction space” era, this framework functioned imperfectly but remained workable because most actions could still be traced back to identifiable human actors and key infrastructure nodes possessed relatively clear territorial affiliations.
The emergence of the intelligent action domain disrupts this assumption of traceable responsibility. AI-driven network behaviors exhibit strong generative and chain-like characteristics. For example, a model on a cloud platform in country A may be invoked using data from country B, deployed via an agent in country C to launch automated attacks against users in country D, ultimately producing consequences on a platform located in country E. The entire action chain spans multiple technological platforms and legal jurisdictions, with key steps generated by intelligent agents. In such cases, the place of action is no longer a single geographic coordinate but a continuously shifting technological pathway.
Likewise, the actor is no longer a single natural person or legal entity but a hybrid generative mechanism composed of human and technological components.
Under these circumstances, traditional jurisdictional principles face serious challenges. Territorial jurisdiction struggles to identify stable connecting points, while personal jurisdiction faces difficulties in identifying the key subjects within a dispersed chain of responsibility. Protective or universal jurisdiction, meanwhile, risks producing overlapping claims among states. This may lead both to “overlapping enforcement” where multiple countries simultaneously assert jurisdiction, and to certain AI-generated acts falling into a “jurisdictional void” where no one claims authority. Global cyberspace governance thus finds itself in a structural dilemma at the level of the distribution of powers and responsibilities.
When it comes to temporal transformation, behavioural iteration and rule supply, traditional global cyberspace governance rules rely mainly on three mechanisms: behavioural norms and declarations formed by international organisations such as the United Nations, regional-level cybercrime treaties and data protection rules, as well as protocols and self-regulatory norms established by technical communities and platforms.
These mechanisms rest upon an implicit assumption that patterns of online behavior remain relatively stable over a certain period.
The temporal reconstruction brought by the intelligent action domain undermines this assumption. AI systems constantly adjust their output patterns through continuous fine-tuning, online learning, and user interaction. Attack methods iterate rapidly within open-source communities, while platform algorithms are constantly optimised in response to data streams. Behavioural patterns thus display persistent evolutionary dynamics. Rules designed to address existing attack methods or the spread of misinformation often become outdated by the time they are implemented.
Consequently, the challenge of rule supply manifests not only as regulatory lag, but also as the continual transformation of regulatory targets. Whether through intergovernmental negotiations or multi-stakeholder mechanisms, it is difficult to identify stable and effective governance targets within a highly dynamic behavioral spectrum. As a result, the rule system risks entering a cycle of crisis emergence-reactive response-renewed obsolescence, placing global cyberspace governance under the persistent structural pressure of perpetually lagged rule supply.
Implementation issues
The third structural impact appears at the level of governance implementation. During the “information interaction space” era, despite the growing influence of platforms and technical communities, states could still control key online behaviors through administrative licensing, content regulation, access restrictions, and economic sanctions. The overall governance structure resembled a hierarchical model of “international norms-state sovereignty-platform execution.”
In the intelligent action domain, however, the distribution of power has shifted. The key nodes that determine how actions are generated have moved from traditional access control points to model and ecosystem definition. Large model providers shape behavioral boundaries through training data selection, model architectures, and alignment strategies. Open-source communities influence the diffusion of offensive and defensive technologies through frameworks and toolkits. Providers of computing power and cloud services indirectly affect the generation and distribution of global network behaviors through resource allocation. These critical nodes are largely controlled by transnational corporations and distributed technological communities, which are difficult for any single sovereign state to fully regulate.
As a result, the effectiveness of global cyberspace governance experiences a double erosion.
On the one hand, states remain responsible for cybersecurity and platform order but find their traditional governance tools less effective when confronting distributed action chains and model-level logic. On the other hand, actors that truly hold the power to define the ecosystem, such as model developers and open-source communities, lack clear responsibility and accountability mechanisms under existing international legal frameworks.
This leads to a structural mismatch where those with power bear little responsibility, while those with responsibility lack effective power. In the absence of effective international coordination, many states have begun to adopt unilateral measures such as data localisation requirements, algorithm registration systems, and export controls on AI models.
While these measures may strengthen domestic regulatory capacity in the short term, they exacerbate the fragmentation of the technological ecosystem at the global level and undermine the capability of the international community to coordinate responses to cyber threats and platform governance. The evolving power structure of the intelligent action domain thus pushes global cyberspace governance into a predicament where increasing investment meets with declining effectiveness. — Contemporary World Magazine
Cai Cuihong is a professor at Centre for American Studies, Fudan University. Contemporary World Magazine is an international relations and social science journal published in Beijing, China.
