New tool tracks when Chinese government blocks websites like ChatGPT

In the months after OpenAI launched ChatGPT in November 2022, the service began gaining popularity in China, with citizens using it to satirize pro-government figures and to help with homework. Since OpenAI restricted access to Chinese users, local developers created mirror sites to make the service easier to access. But ChatGPT’s boom in China was short-lived. The Chinese government blocked ChatGPT’s domain on March 2, 2023, new research has found.

Historically, tracking exactly when Chinese authorities blocked specific domains has been difficult because researchers had to choose which domains to test. However, according to newly launched platform GFWeb, which admitted Rest of the world exclusive first access, in the same month that the Chinese government first blocked ChatGPT, authorities also blocked dozens of alternative chatbots and websites using ChatGPT technology. Rest of the world It was also discovered that popular machine learning platform Hugging Face was blocked in China months before the company reported issues.

GFWeb is now available to the public for free, and it continually tests millions of websites from both China and outside of China to determine exactly when they are no longer accessible to users in China. It detects which sites are blocked by using the unique filtering behaviors of the Great Firewall. The service is funded primarily by the nonprofit Open Technology Fund and has received research contributions from faculty at the University of British Columbia, the University of Toronto, the University of Chicago, and Stony Brook University.

“This system not only improves our ability to track the timing and scope of censorship events, but also helps identify patterns and changes in the strategies used by the Great Firewall,” said Nguyen Phong Hoang, a computer scientist at the University of British Columbia and the creator of the platform. Rest of the world. “I hope GFWeb will help researchers, policymakers, and the public better understand the changing landscape of censorship in China.”

It was previously unclear when Hugging Face was first blocked in China. In October 2023, the company reported “unfortunately accessibility issues” in the country. In fact, GFWeb data suggests that Huggingface.co was actually blocked on May 7, 2023, several months before the company identified the issue.

Data from GFWeb allows observers to spot long-term trends. For example, it shows that Chinese authorities are particularly concerned about AI tools used to generate content. In addition to websites that appear to use ChatGPT technology, most of the blocked AI websites include tools that help with video and image editing. This includes services like OpenArt and VoiceDub.

This suggests that the Chinese Communist Party is “very sensitive to platforms generating content that are not controlled by the regime. That is the main threat,” said Jeffrey Ding, an assistant professor of political science at George Washington University and a leading expert on China’s technological capabilities. Rest of the world.

“Blocking AI sites may not stop software developers from using VPNs (virtual private networks) to access these tools, but it could make it harder for the average Chinese person to use AI to generate politically charged content, such as a video mocking Chinese Communist Party leaders or a couplet about Chinese corruption,” Ding said.

He pointed to a viral AI-manipulated video that circulated on Western social media platforms last year showing Chinese President Xi Jinping saying flattering things about the American public in fluent English. It contained numerous inaccuracies, and Xi never gave a full speech in English.

Ding said Beijing does not want to see AI-manipulated content on its social media that may misrepresent Chinese leaders.

Rest of the worldAnalysis using this tool has shown a correlation between spikes in censorship activity and significant events, such as the passage of new AI regulations. For example, hundreds of religion-related AI sites, including Biblechat.ai and Church.ai, were blocked this spring—coinciding with a surge in launches of spirituality-related AI apps.

“This system not only enhances our ability to track the timing and scope of censorship events, but also helps identify patterns and changes in the strategies used by the Great Firewall.”

In January 2023, China became one of the first countries in the world to introduce regulations governing deepfake AI technology, and in August 2023, it introduced comprehensive regulations for generative AI. These dates also correspond with the increased blocking of AI domains.

The data suggests that authorities are applying existing censorship protocols to new AI technology, said Jeremy Daum, a senior fellow at the Paul Tsai China Center at Yale Law School.

“The biggest names with the most traffic will be blocked. Beyond that, there seems to be a discovery period, so you see batches of sites that are being censored,” he said Rest of the world“The process is always arcane, with some automation at lower levels, but when blocking happens in batches, it is usually a decision by CAC (Cyberspace Administration of China) officials.”

Popular artificial intelligence (AI) chatbots produced by Chinese companies that must comply with censorship regulations include Baidu’s Ernie Bot, Alibaba’s Tongyi Qianwen and ByteDance’s Doubao, as well as a number of solutions developed by Chinese startups.

Phong was previously the principal investigator for GFWatch, a large nine-month study that provided insight into China’s domain-blocking behavior and paved the way for the design of GFWeb. He has worked on smaller studies on internet censorship in Turkmenistan and parts of the Middle East.

Charlie Smith, who uses a pseudonym for security reasons, is co-founder of GreatFire.org, which offers tools to analyze and bypass Chinese internet censorship. Smith said Phong’s work opens up exciting possibilities for researchers.

“The exact date of censorship is helpful for many reasons. For example, we can determine whether sites are blocked in connection with specific events,” he said Rest of the world“It also helps show how authorities are blocking these sites. Do they block mainly on Mondays? Do they work on weekends? We would be able to identify whether authorities are blocking websites in a certain pattern.”

“GFWeb’s system should encourage more people to initiate tests themselves, which could significantly expand our understanding of internet control in China,” Smith said.