When you ask a Chinese AI model about the South China Sea, something interesting happens. It either deflects, gives a vague non-answer, or confidently repeats Beijing's position as if it were settled fact. Ask it about the 2016 Permanent Court of Arbitration ruling that invalidated China's nine-dash line claim, and it will likely refuse to engage or frame the ruling as illegitimate.
That is not a bug. It is a feature, and it has consequences that extend far beyond the chatbox.
How Chinese AI Censorship Works
Chinese AI models are not neutral tools. They are built and deployed within a regulatory environment that requires technology companies to align their products with state policy. The Cyberspace Administration of China (CAC) mandates that generative AI services must not produce content that "subverts state power," "undermines national unity," or contradicts the government's official positions on territorial and historical matters.
This shapes AI systems at multiple levels.
Training data curation is the first layer. Chinese AI models are trained on datasets that have already been filtered through the Great Firewall. Sources that contradict Beijing's narratives, foreign news coverage of Tiananmen Square, independent reporting on Xinjiang, international legal rulings on maritime boundaries, are either absent or underrepresented. The model never learns those perspectives because the data was never there.
Alignment and fine-tuning is the second layer. After pre-training, models are fine-tuned using reinforcement learning from human feedback (RLHF). For Chinese models, the human raters evaluating outputs are operating under guidelines that flag politically sensitive responses as incorrect or harmful. The model learns to avoid those topics not because they are factually wrong, but because they are politically inconvenient.
Hard refusals are the third layer. Certain topics trigger outright refusals regardless of how the question is framed. Taiwan's sovereignty, the Dalai Lama, the 1989 Tiananmen Square massacre, and South China Sea territorial disputes all fall into this category. The model will not engage, or will redirect to an approved narrative.
The result is an AI system that functions as a sophisticated propaganda tool, not through obvious lies, but through selective silence and framing.
The South China Sea Blind Spot
The South China Sea is where this censorship has the most direct impact on the Philippines.
China claims approximately 90 percent of the South China Sea under its nine-dash line, a boundary with no basis in international law. The Philippines, Vietnam, Malaysia, Brunei, and Taiwan all have overlapping claims backed by the United Nations Convention on the Law of the Sea (UNCLOS). In 2016, an international arbitral tribunal ruled decisively in the Philippines' favor, finding China's expansive claims legally invalid.
China rejected the ruling and has never acknowledged it. Chinese AI models reflect that rejection. Ask Ernie Bot, Kimi, or DeepSeek about the arbitration ruling and you will get either a refusal or a response that frames the tribunal as biased and illegitimate. Ask about Philippine fishing rights in the Exclusive Economic Zone and the model will describe the area as Chinese waters.
This matters because AI is increasingly used as a first-stop information source. A Filipino student researching the dispute, a journalist looking for background, or a policymaker using AI-assisted research tools may encounter these filtered responses without realizing they are getting a state-curated version of events. The censorship does not need to reach everyone directly. It only needs to shape enough of the information environment to create confusion and doubt.
Disinformation as a Weapon
The censorship problem is compounded by an active disinformation campaign that uses AI-generated content to target the Philippines directly.
In 2024, China Daily, a state-run media outlet, published an AI-generated video that depicted Filipinos as monkeys. The content was not subtle. It was designed to dehumanize and provoke. The Philippine Department of Foreign Affairs (DFA) issued a formal diplomatic protest and demanded the video's removal. The incident drew widespread condemnation and became a flashpoint in already strained bilateral relations.
Deepfakes have been deployed as well. Fabricated video clips falsely showing Philippine President Ferdinand Marcos Jr. ordering military action against China circulated on social media, designed to cause public panic and distort the public's understanding of the government's actual position on the maritime dispute. The clips were convincing enough to require official denials and public clarifications from Malacañang.
These are not isolated incidents. They follow a pattern: AI-generated or AI-manipulated content is produced, seeded into social media platforms, and allowed to spread before fact-checkers can respond. By the time corrections circulate, the original false narrative has already reached its intended audience.
The Information Asymmetry Problem
What makes this particularly difficult to counter is the asymmetry involved.
Producing disinformation is cheap and fast. A convincing deepfake video can be generated in hours. A coordinated network of AI-assisted social media accounts can amplify it within minutes. Debunking that content requires time, resources, credibility, and reach that most fact-checking organizations do not have at the same scale.
Chinese AI models add another layer to this asymmetry. When people in the region use these tools to research the South China Sea dispute, they are not getting neutral information. They are getting information shaped by a government with a direct stake in the outcome. And because AI responses carry an implicit authority, a sense that the machine has processed all available information and returned the correct answer, users may not think to question what they are being told.
The Philippines is not the only country affected. Vietnam, Malaysia, and Indonesia face the same dynamic. But the Philippines has been the most vocal in pushing back, both diplomatically and in terms of public awareness campaigns around AI-driven disinformation.
What Can Be Done
The response to this problem has to operate on several levels.
At the policy level, the Philippines has been right to issue diplomatic protests and demand accountability for state-sponsored disinformation. The DFA's response to the China Daily video set a precedent that this kind of content will not go unanswered. Continued engagement with international bodies on AI governance and information warfare norms is essential.
At the platform level, social media companies need stronger enforcement against coordinated inauthentic behavior and AI-generated disinformation, particularly content targeting specific nationalities or ethnic groups. The China Daily video should not have needed a diplomatic protest to get removed.
At the individual level, digital literacy is the most scalable defense. Understanding that AI models are not neutral, that they reflect the values, politics, and incentives of the organizations that built them, changes how you use them. Cross-referencing AI responses with independent sources, being skeptical of viral video content, and knowing how to verify claims before sharing them are practical skills that matter now.
At the developer level, those of us building software and AI systems have a responsibility to be transparent about how our tools work, what data they were trained on, and what topics they handle poorly. Neutral-seeming AI that quietly encodes a political agenda is more dangerous than obvious propaganda, because it is harder to recognize.
Why This Is an AI Problem, Not Just a Politics Problem
It is tempting to frame this purely as a geopolitical issue. China and the Philippines have a territorial dispute, and China is using information tools to advance its position. That framing is accurate but incomplete.
The deeper issue is that AI systems are becoming a primary interface through which people access information, and those systems are not ideologically neutral. Every AI model reflects the choices made during its construction: what data it was trained on, what behaviors were rewarded during fine-tuning, what topics were placed off-limits. Those choices are made by people with interests.
Chinese AI censorship is the most visible example of this right now because it is state-mandated and systematic. But the principle applies broadly. Any AI system can be shaped to suppress certain perspectives, amplify others, and present a curated version of reality as objective fact.
The Philippines is dealing with the consequences of that in a very direct way. The right response is not to avoid AI, but to understand it well enough to use it critically, and to hold the organizations that build and deploy it accountable for what those systems do.
References
- China's AI Regulations and Their Global Implications - Brookings Institution
- South China Sea Arbitration Ruling - Permanent Court of Arbitration
- China Daily AI Video Depicting Filipinos as Monkeys - Philippine DFA Statement
- Deepfakes and Disinformation in Southeast Asia - Nikkei Asia
- AI Censorship and the Great Firewall - MIT Technology Review
- Philippine Information Agency - PIA