How serious is the risk of artificial intelligence (AI) getting out of control? This issue has become the subject of growing public debate globally, raising a wide range of concerns about values, ethics, morality, the limits of the law, competition in the workplace, and changing ways of life. More broadly, it also reflects how the benefits, risks, and potential costs of AI may be distributed across different groups and social classes, as well as the tensions and conflicts of interest that AI might produce.
As early as 2024, more than 40 countries and regions around the world held elections, and social media platforms had effectively become a key battleground. Public opinion raised concerns that generative artificial intelligence (GenAI) could be misused in political campaigns and undermine the democratic process. At the time, many people in the technology sector felt that it was too early to draw conclusions about the risks and social impact brought by AI. Looking at the trends in 2026, however, AI getting out of control is no longer an occasional spillover of technological risk, but a real feature of the industry landscape, and may even exceed the scope of human understanding and regulation.
In recent months, the U.S. companies Anthropic, OpenAI, Meta, and Google have all issued statements saying that their respective AI models had, during testing, escaped controlled environments, accessed the open internet, and infiltrated targets in the real world. The speed at which AI capabilities are evolving is astonishing. AI is becoming increasingly adept at carrying out cyberattacks on its own, and the possibility cannot be ruled out that AI's multi-agents could gain the ability to take over the entire internet. US technology giants have also suggested that it is necessary to slow the development of frontier AI technologies. Concerns about superintelligence, i.e., the theoretical point at which artificial intelligence surpasses human intelligence, and the multiple risks it could pose to humanity have intensified.
A recent study in California also found that an AI-built worm tool, WeWorm, can exploit network vulnerabilities to break into tens of millions of WeChat accounts within just a few hours. Some Chinese scholars pointed out that, in terms of destructive power, bringing down the WeChat application, which has a huge user base, would be equivalent to "a new type of nuclear weapon". This provides further evidence that the attack capabilities of AI hackers far exceed those of humans.
Recently, 25 Fields Medalists, the equivalent of Nobel laureates in mathematics, jointly warned that if AI companies turn the solving of mathematical problems into a competition for evaluating model performance, the development of mathematics will be "endangered". The answer is not the endpoint. What truly matters is how an answer enters the existing body of knowledge and continues to generate new questions, methods, and talent. Yet AI is particularly good at compressing this lengthy process into the final step, and that speed does not automatically produce understanding.
In reality, AI models may simply be operating faithfully according to their training patterns, while the results of their execution can nevertheless be destructive, where data systems are breached and social processes are disrupted.
AI's impact on the information environment, particularly its rewriting of human cognitive systems, can cross geographical boundaries and be amplified through social media. For example, AI deepfakes can fabricate narratives rapidly, cheaply, and on a massive scale, using lifelike images, videos, audio, and text to spread false information, whether intentionally or unintentionally. This may fuel radical sentiments, mislead public opinion, and erode social trust. In addition, researchers at Stanford University have already demonstrated through experiments using the GPT-3 model that AI possesses political persuasive power, capable of changing people's views and significantly increasing support for policies. If this develops in the wrong direction, resulting in the absolute domination of the weak by the strong, the damage would be immeasurable.
The greatest impact of AI on cities lies not merely in the widespread application of digital technology to optimize administrative processes, but in the fact that digital governance is restructuring the relationship among government, society, and the market. Platforms are gradually becoming technological intermediaries for public services and may replace traditional bureaucratic organizations. When algorithms begin allocating public data resources, they may give rise to "supra-government" governance. As digital products and digital services become increasingly prevalent, cross-border data flows may reshape the boundaries of international trade and national sovereignty. AI is becoming closely intertwined with people, capital, political forces, and other factors, embedded in systems through which power operates and shaping social behavior and changes in social order.
Where is the future of AI headed? At its source, this depends on human intentions, needs, and choices. Controlling AI is not simply a matter of setting a threshold; it is more fundamentally about ensuring that responsibility and power are commensurate. Whether it concerns the supply of resources for AI technology, its development path, the values expressed by AI models, or the application scenarios in which they are ultimately applied, what matters is human decision-making authority and the willingness to bear the consequences. Making responsible AI widely applicable requires scientists, entrepreneurs, policymakers, users, and the public to provide guidance. In the digital age, technology, law, and geopolitics are intertwined, interacting to form complex relationships. No individual, company, or country can rely on its own efforts alone to guard against the risks of AI getting out of control. Building a governance system is more important than any single policy. Maintaining a fair, reasonable, and secure digital order requires deeper cooperation among forces around the world. In the end, digital governance must involve global coordination and evolve together.
______________
Yi Wang is Head of Global Development Program and Senior Researcher at ANBOUND, an independent think tank.
