Chinese response: Xi Jinping's information control for the first time got "training data export"

In the past, the understanding of the CCP network censorship has a clear geographical boundary.

People in China, search engines do not see, social platforms remove political sensitive content, news agencies receive propaganda instructions; after leaving the Chinese network firewall, at least in principle can re-contact the information being blocked.

Generated artificial intelligence is breaking this old boundary.

The Wall Street Journal’s August 13 survey tested the performance of U.S. AI systems such as ChatGPT, Claude and Gemini dealing with sensitive Chinese political issues. The report found that in some Chinese issues involving Chinese political and authoritative leaders, models could appear similar to China’s censored chat bots, self-censorship or narrative deviations. The survey pointed one reason to training data: the model-learning Chinese Internet itself has long been screened by China’s censorship and propaganda system.

The danger of this is not that a model is occasionally answering the question of Xi Jinping. Large language models would have produced factual errors, and there are huge differences in the quality of data from different language training. What is really worth warning is a structural possibility: when a language's large-scale public information has been deleted and rewritten by the country for a long time, global AI absorbs this data as a sample of human knowledge, and censorship may upgrade from "deleting information" to "the raw material for shaping the world of machines".

The Chinese Communist Party does not need to own an American AI company, nor does it need to order a Silicon Valley enterprise to delete it.As long as the political information that can survive on the Chinese network itself has been seriously imbalanced, the language library facing the training model naturally carries this imbalance.

This is an unprecedented promotion advantage in the past.

Traditional propaganda requires audiences to watch Xinhua, Central TV or the People's Daily. Algorithmic pollution does not require audiences to know where information comes from. A Dutch student, an American researcher or a Chinese overseas only need to ask AI in Chinese, the answers may have inherited the structural gaps left in China's Internet censorship for years in the probability model.

Of course, this cannot be exaggerated as “American AI has been controlled by the Chinese Communist Party.” The Wall Street Journal tests revealed deviation risks worth systematic research, rather than proving that Beijing can directly manipulate these models.

But precisely for this reason, the solution cannot only require a model of "anti-CPC".

The truly reliable technical answers should be training data transparency, diversification of sources, cross-lingual fact calibration, and giving models access to unchecked Chinese historical archives, academic research, international media and first-hand testimonies on politically sensitive issues.

The Xi Jinping administration has spent years investing enormous resources in controlling what Chinese people can see.The new problem that emerges today is that after a decade, controlled information environments are becoming the learning material for the next generation of artificial intelligence.

In the past, the firewall prevented Chinese from seeing the world.

If global AI companies do not carefully deal with Chinese training data pollution, the next legacy it leaves is likely to let the world’s machines learn to understand China in a way within a firewall.

Artificial intelligence # network censorship # Xi Jinping # Communist Party propaganda # information war # viewpoint comments