From ChatGPT to Chat CCP: The Future of Generative AI Models in China

Only three months into its release, ChatGPT has already created a frenzy among people around the world. Source: Reuters

Chinese regulators recently warned tech companies not to release ChatGPT-like products, for fear that chatbots would spread “disinformation” by the U.S. government. With all the hype surrounding OpenAI’s natural language model setting off an “AI Arms Race,” does this clampdown set the People’s Republic of China (PRC) back in the global technological competition? 

Not necessarily. Instead, Chinese tech giants will likely develop their own versions of ChatGPT under Chinese Communist Party (CCP) oversight.

Silicon Valley-based OpenAI launched ChatGPT on November 30, 2022, and it reached more than 100 million users within the first two months. Not only could the large language model (LLM)-trained chatbot provide accurate answers to human-generated questions, but it could also respond in prose, as if users were actually “chatting” with a real person. As Microsoft poured even more money into OpenAI, Google, Anthropic, and even Buzzfeed jumped at the chance to invest in future generative AI projects. Chinese tech giants Baidu, Alibaba, and JD.com announced shortly thereafter that they, too, would release generative AI models that could rival ChatGPT.

Yet on February 22, 2023, regulators in the PRC told Alibaba and Tencent that they could not offer access to ChatGPT services, and that any companies who wish to launch similar products must notify government officials beforehand. Though ChatGPT is currently accessible in China only through a virtual private network (VPN), this announcement induced Tencent to immediately suspend services by third-party developers that claimed to provide access to ChatGPT and similar “copycats.”

ChatGPT and similar tools have caused a headache for the PRC’s authoritarian state because their responses to human-inputted questions have ventured into sensitive political topics. For example, when asked about the state of China’s economy, a chatbot created by Chinese AI company Yuan Yu provided a list of China’s economic problems; another prompt led the chatbot, called ChatYuan, to name Russia as an aggressor in its war on Ukraine. These answers are problematic in the PRC because the government typically censors criticism of its economic affairs and it has likewise been very cautious not to officially take sides in Putin’s war. ChatYuan was shut down within three days of its release.

These “problems” with generative AI tools like ChatGPT have arisen from the nature of how LLMs are trained. LLMs are deep learning algorithms that recognize, predict and generate text (or other content) based on knowledge acquired from data. Through unsupervised learning – an AI training method that does not provide algorithms instructions for how to interact with data – LLMs train on massive datasets and “learn” how to interpret datasets on their own. Thus, since ChatGPT and similar algorithms have been trained on data that has not been censored by the CCP (mostly sourced from Western countries), their answers to questions may stray from the Party’s narrative. There have also been instances where AI-generated responses in Simplified Chinese (used in mainland China) differed from those in Traditional Chinese (used outside of mainland China); answers in the former tended to mirror CCP propaganda, while answers in the latter were more open and honest about sensitive political topics.

In order to continue cultivating indigenous technological capabilities and stay on par with U.S. innovation, the CCP is likely to lift its ban on ChatGPT-like services once it derives a set of rules and regulations that safeguard its interests. This might include increased scrutiny over the code used in algorithms, requirements that AI models train on Party-approved datasets, a rigorous testing period for authorities to approve answers to sensitive queries, and severe punishments for developers whose chatbots produce controversial responses. Eventually, LLMs might join the likes of state-owned newspapers as mouthpieces for the government, because the information from Party-censored data would be all they know.

What does this mean for the future of generative AI models in China? For one, the CCP’s efforts to restrict free speech in chatbots is yet another step toward developing a global “splinternet” – an Internet fragmented by, in this case, governments in order to hinder the free flow of information across borders. The PRC’s “Great Firewall,” which already shields the Internet accessed by Chinese citizens from the rest of the world, has already begun this process, while the Kremlin’s recent move to isolate and restrict traffic moving across the Russian web has accelerated its creation.

Furthermore, the CCP’s swift crackdown on AI-powered chatbots may become a model to other autocratic governments around the world that are anxious about what LLMs might do in their respective countries. The PRC is already exporting its surveillance technology to governments seeking tighter control over adversaries; it is also seeking to spread its governance model through economic incentives and coercive measures. Beijing’s decision to censure ChatGPT and similar tools will surely send a warning signal to leaders who look up to the Chinese government. The PRC’s own version of an LLM may very well be the next “exported” technology seeking to shape global narratives around the CCP’s messaging – and inspire other countries to create their own. 

While the instinct of many in the West might be to criticize such restrictions on private companies as a hindrance to innovation, the PRC technological ecosystem may indeed surprise us. Just as the Chinese government has carefully orchestrated its post-World Trade Organization admittance “market economy” into one that some describe as “party-state capitalism,” and likewise deployed the world’s most sophisticated censorship regime on an Internet that many once believed would bring democracy to China, the CCP can very well leverage AI to achieve its goals, too. We have seen this movie before, and we know how it will end.