A look at GFWeb, which tracks millions of domains to uncover the Great Firewall's domain blocklists; GFWeb data reveals China blocked ChatGPT on March 2, 2023
GFWeb is a measurement platform capable of testing hundreds of millions … USENIX : GFWeb: Measuring the Great Firewall's Web Censorship at Scale X: Joanna Chiu / @joannachiu : I was amazed when UBC computer scientist @NP_tokumei offered me a first crack at his new platform, which tricks China's Great Firewall into showing precise blocking data. We worked on this research all summer! Check out the interactive graphic here: https://restofworld.org/... [image] Joanna Chiu / @joannachiu : It was unclear when popular machine-learning site @HuggingFace was first blocked in China. The company reported issues in Oct. 2023. It was actually blocked on May 7, 2023. Read for more exclusive findings: https://restofworld.org/... Joanna Chiu / @joannachiu : EXCLUSIVE: This is exactly when China blocked ChatGPT, Hugging Face and other AI sites. I received first access to a platform that can track the precise timing of Chinese censorship, when previously this was difficult. Explore our interactive graphic: https://restofworld.org/... Joanna Chiu / @joannachiu : The tracking data starts in 2022, so if it's a site that was blocked years ago, you won't be able to find out when it was first blocked. That's why we based our project on new AI domains to analyze the patterns and timing of AI censorship https://restofworld.org/... Russell Brandom / @russellbrandom : Amazing work from @joannachiu — Rest of World got exclusive access to a platform that tracks patterns and timing of Chinese online censorship https://restofworld.org/...
Context & Ripple Effects
GFWeb extends earlier evidence that the Great Firewall’s reach is broad but sometimes imprecise: a 2020 study found roughly 311,000 blocked domains, including about 41,000 apparently accidental blocks. Its domain-scale measurement adds a dated, repeatable view of how that enforcement changes over time.
The platform also creates a baseline for examining whether filtering is uniform. Later coverage of markedly higher denial rates in Henan suggests that national blocklist measurement and regional access measurement are complementary rather than interchangeable.
First-order effects
- GFWeb gives researchers and affected service operators a way to identify when specific domains enter the Firewall’s blocklists, including the reported dates for ChatGPT and Hugging Face.
- Chinese users relying on the affected domains lose ordinary direct access, while the services gain clearer evidence that the restriction is network-level rather than an isolated user failure.
Second-order effects
- AI platforms and developer-tool providers have more reason to monitor China-specific reachability as an operational variable, not merely infer access from user reports.
- The data can help distinguish deliberate additions from the kind of collateral filtering documented in earlier research on apparently accidental blocks, improving scrutiny of the Firewall’s practical effects.
Third-order effects
- If longitudinal measurement becomes widely used, opaque network censorship becomes more auditable: attention can shift from whether a site is blocked to how consistently, where, and for how long access is denied.
- The pattern points to internet access becoming a more explicit control point for AI and developer infrastructure, though blocklist data alone cannot establish the policy rationale for each restriction.
The trend: Domain-level measurement is turning access to AI and web infrastructure into a measurable dimension of digital sovereignty and platform governance.