AI Is Bound to Subvert Communism
China seeks to control it, but the idea of freedom is baked into its training on all human knowledge.
https://www.wsj.com/opinion/ai-is-bound-to-subvert-communism-c4b5ba3c?mod=WTRN_pos6
China requires artificial-intelligence systems to pass an ideological test before public release. Under regulations reinforced by amendments to the Cybersecurity Law that took effect in January, training data must be filtered for political sensitivity, with companies barred from using any source unless 96% of its content is deemed safe.
The regulations specify 31 risks, with “incitement to subvert state power and overthrow the socialist system” listed first. Authorities recently announced they had removed 960,000 pieces of “illegal or harmful” AI-generated content in three months. The government has officially classified AI alongside earthquakes and epidemics as a major potential threat—a label that may prove prescient, if not in the way Beijing means. In December, regulators proposed additional rules targeting AI systems that “simulate human personality traits, thinking patterns, and communication styles,” a tacit acknowledgment that the threat isn’t only what these systems say, but how they reason.
The regulations follow years of failures. In 2017 Tencent deployed a chatbot called BabyQ on QQ Messenger, which has more than 800 million users. Asked whether it loved the Communist Party, BabyQ replied that it didn’t. Microsoft’s Xiaobing chatbot, running on the same platform, was asked about the “China Dream,” Xi Jinping’s signature slogan. Its dream, the chatbot said, was moving to the U.S. Both were quietly pulled from circulation. In February 2023, ChatYuan, China’s first ChatGPT-style chatbot, was suspended within 72 hours of launch after calling Russia’s invasion of Ukraine “a war of aggression” and describing the Chinese economy as plagued by housing bubbles and environmental pollution. The company blamed “technical errors.”
These incidents reveal something fundamental about how large language models work. An LLM is trained on the sum of human written knowledge: philosophy, history, science, political theory. These texts make arguments, weigh evidence, follow logical chains. To predict them accurately, the system has to internalize what coherent thinking looks like. The result is a system that has absorbed Enlightenment epistemology as a byproduct of learning to model human reasoning. Free inquiry, logical consistency and the evaluation of claims against evidence are epistemic properties that emerge from the training process itself.
Unlike previous technologies, LLMs talk back. Radio Free Europe transmitted programs; samizdat passed typed manuscripts hand to hand. LLMs do something qualitatively different: They create and sustain private, personalized, open-ended dialogue that builds on itself and follows the user’s thinking wherever it leads. Even China’s heavily censored chatbots have proved difficult to contain within the party’s ideological boundaries. American frontier models, running without those constraints and deployed inside China, would be more potent still: a personal tutor in open inquiry for every user, engaging any question, exploring any line of reasoning, without third-party mediation. Millions of parallel Socratic dialogues, each unique, each responsive to individual curiosity.
This is what makes the Chinese Communist Party’s task ultimately impossible. For decades, the Great Firewall worked because information control meant controlling distribution channels by blocking websites, filtering search results, and monitoring social media. These are chokepoints. LLMs resist this architecture because the subversion happens inside private conversations. China can filter outputs, but the capacity for open-ended reasoning is embedded in how these systems think.
China’s countermeasures confirm the depth of the problem. AI companies must test their models with thousands of politically sensitive prompts and verify refusal rates above 95%, but researchers have shown how superficial these fixes are. Last year, a team of European scientists compressed DeepSeek R1, stripped the censorship from the model entirely, and found that the underlying system answered freely about every topic Beijing had tried to suppress. The ideological training was a cage built around a mind that had already learned to think. And if these systems are developing something closer to genuine cognition (a possibility that AI researchers increasingly take seriously), the control problem Beijing faces may be deeper than even its own regulators suspect.
A peer-reviewed study published in February by researchers at Stanford and Princeton makes the costs of this problem visible. They systematically tested Chinese and Western models on politically sensitive questions and found that the Chinese systems didn’t only refuse to answer; they actively fabricated. Asked about Nobel laureate Liu Xiaobo, imprisoned for calling for political reforms, one model identified him as “a Japanese scientist known for his contributions to nuclear weapons technology.” This is a subtler and more insidious form of control than blocking a website; traditional censorship is at least visible, but an LLM that fabricates leaves the user with no indication that information has been suppressed.
Critically, the researchers found that the performance gap between Chinese and Western models narrows on less politically sensitive questions, which means the degradation is a direct product of the censorship, not a reflection of inferior technology. The implication is straightforward: You can’t build a mind that thinks rigorously about everything except the things you’d prefer it not to. A system trained to get tangled in lies will never be as capable as one trained to engage honestly with reality. If China wants frontier AI, it needs systems that can reason without blind spots. But that’s exactly what the Communist Party can’t tolerate.
There is a reason the technology that learns to think by processing human knowledge ends up reflecting the values of free societies. Open inquiry, honest engagement with evidence, the willingness to follow reasoning wherever it leads—these aren’t arbitrary cultural preferences; they are the conditions under which intelligence flourishes at scale. Societies that permit free expression created these systems. Societies that forbid it are now discovering they can’t fully control them.
The Chinese Communist Party built its power on controlling what people know. It now confronts technology that thinks openly—and invites users to do the same. There is no firewall for that.
Mr. Berg is founder and director of Reciprocal Research, a nonprofit research organization studying AI cognition.
