A disclosure from Anthropic this September revealed something digital rights researchers had long warned about: generative AI is now being actively used to build surveillance dossiers on dissidents, religious leaders, and diaspora activists. The finding, paired with Hong Kong's announcement of a sweeping new "AI City Brain," marks a turning point in how artificial intelligence intersects with state power in and around China.
An AI Model Turned Intelligence Tool
Anthropic's report, Detecting and Countering Misuse of AI, disclosed that a suspected China-based, government-aligned operation misused its Claude models between December 2025 and August 2026. The operators did not use the AI to break into systems or steal data through conventional hacking. Instead, they directed Claude to perform the analytical labor of intelligence work: drafting personnel profiles, compiling "clue reports," and generating daily situational briefings on individuals and organizations that Chinese authorities consider politically sensitive.
The targets were wide-ranging and specific. They included senior Catholic cardinals across Asia, leaders of the Presbyterian Church in Taiwan, the Central Tibetan Administration, advocacy groups such as the International Campaign for Tibet and Students for a Free Tibet, Falun Gong practitioners and affiliated outlets, and organizers of Tiananmen Square commemorations. Operators reportedly gathered birth dates, immigration records, and social media footprints, and conducted reconnaissance on religious buildings, including floor plans and structural diagrams. Monitoring spanned both domestic Chinese platforms such as WeChat and Weibo and international ones including LinkedIn, Instagram, and X.
What makes this case significant is not novelty but efficiency. Compiling the kind of dossier described in the report once required teams of analysts working for days or weeks. An AI model can now produce a comparable synthesis in a fraction of the time, lowering the operational cost of surveillance at scale. Anthropic said it found no evidence the operators accessed other users' private conversations, but the episode follows an earlier disclosure by OpenAI, which found that an individual linked to Chinese law enforcement had used ChatGPT to edit internal reports on operations against dissidents and foreign political figures. Together, these cases point to a pattern: state-linked actors treating commercial AI platforms as productivity tools for political surveillance, not merely as novelties or chatbots.
The implications extend beyond any single company's moderation policies. AI developers now function, whether they intend to or not, as a line of defense against state misuse of their own products. That responsibility includes detecting abuse patterns, disclosing them publicly, and building safeguards into how models handle requests involving named individuals or organizations. For human rights defenders and advocacy groups, the lesson is more immediate: information shared with AI tools, even in seemingly routine interactions, can potentially be aggregated and exploited if platforms are compromised or misused.
Hong Kong's "AI City Brain" and the Oversight Gap
Ten days after Anthropic's disclosure, Hong Kong Chief Executive John Lee announced in his annual Policy Address plans to build an "AI City Brain" - a system designed to integrate and analyze data from government departments and digital infrastructure in real time. Officials describe the goal as improved urban management and faster emergency response, and a senior official has since said the system is not intended to surveil individuals.
The design nonetheless raises hard questions. The proposal folds in law-enforcement infrastructure, including the Security Bureau's Emergency Monitoring and Support Centre and a police-led Inter-departmental Major Incident Platform. Combining these systems with a city-wide data analysis engine creates a capability that functions very differently from isolated departmental databases. When information from transport, utilities, emergency services, and policing systems sits inside a single AI-driven platform, the ability to cross-reference and infer behavior patterns about specific individuals expands considerably, regardless of the system's stated purpose.
There are structural risks as well. Centralizing large volumes of personal data increases the potential damage from any single breach or unauthorized access. AI-generated assessments can also contain errors or embedded bias, and such outputs carry real consequences if used to inform official decisions about individuals. None of these risks are unique to Hong Kong, but they are sharpened by the territory's political climate, where independent oversight mechanisms and avenues for public accountability have narrowed considerably in recent years. Without clear limits on data access, retention periods, and external review, a system billed as administrative efficiency can quietly become an instrument of comprehensive monitoring.
A Broader Pattern of AI-Enabled Control
These two developments sit alongside a wider set of trends documented this month: expanding censorship campaigns using AI-assisted content detection, propaganda systems that simulate foreign audiences to refine messaging before release, and a leaked moderation lexicon showing how platforms like Douyin and Toutiao already screen thousands of sensitive terms. Taken together, they suggest that AI is being woven into state information control at multiple levels simultaneously - from the individual dossier to the city-wide dashboard to the national content filter.
For researchers, journalists, and ordinary users navigating these systems, the practical takeaway is caution rather than alarm. Encryption, careful platform selection, and skepticism about what data any AI tool retains remain the most reliable defenses available, even as the surveillance capabilities arrayed against vulnerable communities grow more sophisticated.