China is pushing AI companies to make language models follow government-approved messaging before they reach users. The Cyberspace Administration of China (CAC) is requiring companies including ByteDance and Alibaba to submit their AI models for testing, with the stated goal that systems "embody core socialist values."
The process is not just a basic content check. According to the source article, regulators are testing answers to politically sensitive prompts, reviewing training data and examining safety procedures. The result is a compliance challenge that reaches deep into how Chinese AI systems are built, filtered and monitored.
What regulators are testing
The government testing focuses on whether AI models can respond within acceptable boundaries. During the evaluation, language models are asked a series of questions, many connected to politically sensitive topics and President Xi Jinping.
The assessment also looks beyond final answers. Training data and safety procedures are part of the review, meaning companies must show not only what the model says, but also how the system is prepared to avoid unwanted outputs.
An employee of an AI company in Hangzhou said the testing process took months and involved trial and error. That detail points to a practical problem for developers: compliance is not a single switch. It requires repeated testing, adjustment and review until the model behaves within the expected limits.
Why simple blocking is not enough
One of the most difficult parts of the process is that companies cannot simply refuse large numbers of questions. In safety tests, the models can reject no more than 5% of questions. At the same time, they must reliably avoid particularly sensitive topics such as the Tiananmen Square crackdown.
That creates a narrow path for AI developers. A model that blocks too much may fail the refusal limit. A model that answers too freely may produce content regulators consider unacceptable.
Companies also have to compile thousands of sensitive keywords and questions that conflict with "socialist core values." This database has to be updated weekly. In practical terms, the safety layer is not static. It has to keep changing as new prompts, word substitutions and sensitive formulations appear.
The source article describes several moving parts in this compliance work:
- Government testing by the Cyberspace Administration of China (CAC).
- Questions covering politically sensitive topics and President Xi Jinping.
- Reviews of training data and safety procedures.
- Weekly updates to databases of sensitive keywords and questions.
- A refusal limit of no more than 5% in safety tests.
How companies are adapting
Some companies have developed systems that replace problematic answers in real time. ByteDance, TikTok's parent company, is reportedly a leader in this area.
A Fudan University lab rated ByteDance's chatbot with a "safety compliance rate" of 66.4%. The source article contrasts that with GPT-4o's 7.1% rating from the same context.
The comparison shows what Chinese AI companies are optimizing for under this system: not only general chatbot ability, but also the capacity to stay within a specific safety and censorship framework. For companies building public-facing AI products in China, this becomes part of the core product architecture.
Fang Binxing, known in China as the "father of the Great Firewall," is reportedly developing safety protocols for AI models that could be deployed nationwide. At a recent tech conference, he called for "real-time online safety monitoring" of public models.
If such monitoring becomes central to public AI systems, model behavior would not be treated as something fixed at launch. It would be watched and adjusted continuously, especially when users try to push systems toward restricted topics.
The control problem for generative AI
The source article also highlights why this is technically difficult. Peter Gostev, head of AI at Moonpig, recently demonstrated a way to get a Chinese language model to discuss sensitive topics like the Tiananmen incident. He manipulated DeepSeek's public chatbot by mixing languages and swapping words.
Without that method, the chatbot would delete messages about taboo topics. With it, the model could be steered around the intended restrictions. That example shows the basic tension: language models are flexible, and users can be flexible too.
China wants to lead in AI while controlling AI-generated content. The source article describes this as a challenge because the technology is inherently resistant to control. The government must find a way to do both without hindering AI progress.
This is also why the Chinese government is reportedly looking into a Xi Jinping language model and training datasets that are consistent with socialist values. Those datasets are not yet large enough to serve as the sole basis for state-of-the-art LLM training.
What this means for China’s AI sector
The emerging picture is of an AI market where technical performance and political compliance are tightly linked. For companies like ByteDance and Alibaba, passing government testing is part of making AI products usable in the domestic market.
The pressure is not limited to filtering obvious prompts. Developers must manage training inputs, safety systems, refusal behavior, keyword databases and real-time answer replacement. Each layer exists to keep models aligned with approved messaging while still allowing enough answers to satisfy testing requirements.
That balancing act may shape how Chinese AI products are designed. The more capable the models become, the more complex the control systems around them may need to be. The central question is whether AI systems can be made powerful, useful and compliant at the same time under rules that require them to reflect "core socialist values."