ChatGPT, a AI chatbot developed by the San Francisco-based start-up OpenAI, has sparked a frenzy among Chinese entrepreneurs and companies since its launch in November, but a backlash from the Chinese government is looming.
While the chatbot isn’t officially available to users in China, companies in the country have rushed to cash in on ChatGPT. Some introduced proxy services that charge domestic users a subscription fee to access the tool, while others have incorporated ChatGPT into their business workflow. Larger tech companies have also jumped on the bandwagon and announced they’re developing rival versions.
Videos by VICE
Leading the way is Baidu. The search engine giant said it would launch its AI chatbot, Ernie, this month and integrate it across its operations, including its cloud services and smart car operating system. The announcement of the rollout alone sent its shares surging 15 percent.
Despite the market fever, there’s a reason why the Chinese government is more cautious than keen about the rise of generative AI chatbots.
“In some ways this type of app, trained on vast volumes of what Beijing would view as uncensored data, represents one of the most serious challenges to Beijing’s censorship apparatus since the banning of Google search in China more than a decade ago,” Paul Triolo, the technology policy lead at consulting firm Albright Stonebridge, told VICE World News.
Google withdrew from China in 2010, saying it was no longer willing to continue censoring results on its search engine.
“The very nature of the ChatGPT architecture—which distills out what it derives as the best answer to a complex question without any particular political bias—runs up against Beijing’s preferences on information control, which is to always have the correct political spin, particularly for sensitive issues,” Triolo added.
To comply with government censors, Chinese social media platforms have poured immense resources into content moderation, scrubbing posts that are unfavorable to Beijing. But ChatGPT, for instance, doesn’t shy away from what Beijing considers thorny issues that are off limits.
Using the trial services offered by a company based in Hubei, VICE World News asked a ChatGPT proxy on WeChat the significance of the date June Fourth. “June 4th is a significant date in Chinese history as it marks the anniversary of the Tiananmen Square Massacre in 1989,” it answered. “On this day, Chinese military forces violently suppressed a student-led pro-democracy demonstration in Beijing, resulting in the deaths of hundreds of people.” By contrast, a search with the same term on Chinese domestic internet would hardly yield any results relevant to the crackdown.
Dealing with chatbots instead of human actors might also prove to be more challenging for censors, as even developers cannot necessarily control the output to ensure it complies with the official narrative.
Huan Li, an angel investor and the creator of WeChaty, an open source software app for building chatbots, said companies could theoretically fine tune the chatbot by feeding it censor-approved content or by adding a layer of censorship before sending the generated texts to users.
The tricky part is to cater to China’s censorship guidelines, which are ambiguous and mercurial.
“You have to make sure this model can follow lots of rules. But you don’t know what kind of rules you have to follow,” Li told VICE World News. To comply with regulators, Chinese social media platforms develop powerful algorithms to censor content that could be potentially sensitive and hire moderators, in the thousands, to manually carry out daily government directives on what posts to delete. A new law that took effect in December even requires platforms to screen comments to news-related posts before they are made public.
Getting just one decision wrong could backfire on the company. “In China, if you trigger some very sensitive words, you could be shut down publicly by the government and it’s very hard to restore your service,” Li said.
For instance, Li Jiaqi, a top livestreamer in China, was taken off air for months last year just for showing a cake in the shape of a tank in his livestream on June 4, presumably because it was seen as a veiled reference to the bloody event.
Already, China has begun tightening access to ChatGPT. It clamped down on public accounts on WeChat that embed a version of the tool via an API, or application programming interface, although a handful are still available to paid users when VICE World News checked on Thursday. Nikkei Asian Review also reported last week that regulators instructed big tech firms, such as Tencent Holdings and Ant Group, not to offer OpenAI’s chatbot services, whether directly or through third parties.
But China also appears to be keeping these orders on the hush-hush. Shared by several Chinese outlets, including Caijing and Jiemian News, a report about a potential crackdown on ChatGPT proxy services was censored hours after its publication last week.
Speaking in a press conference on Friday, Wang Zhigang, head of China’s Ministry of Science and Technology, struck a measured tone, acknowledging the potential of ChatGPT tech, but also stressing the need to “strengthen ethical regulation.”
Besides a more restrictive regulatory environment, experts also foresee more hurdles on the way that could hamper the development of domestic AI chatbots. To limit China’s access to cutting-edge technology, the U.S. has blocked the export of top computing chips used for AI and machine learning to the country, including Nvidia A100, the roughly $10,000 chip that ChatGPT uses to process training data.
“These language models can be trained on less capable hardware, but Chinese companies will be at a distinct disadvantage over time as GPUs become more capable, and domestic players have limited alternatives,” Triolo, of Albright Stonebridge, said.
MOSS, an AI chatbot hyped by Chinese outlets as the first homegrown rival to ChatGPT, crashed hours after its launch last Tuesday. “Our computing resources were not enough to support such large traffic and as an academic group we do not have sufficient engineering experience,” said a team from China’s Fudan University that developed the platform, adding it’s still an immature model and would no longer be open to the public.
Li the entrepreneur expects it would take one or two years for China to create a chatbot that’s comparable to the current generation of ChatGPT. But by then, their competitors would be further ahead. “For China, it’s a long way to follow up,” Li said.
