Great Firewall of China: Technical History, Censorship Mechanisms, Cases, Data and Circumvention

Living research project • expanded public edition: September 18, 2026

The Great Firewall of China (GFW) is the commonly used name for the network-level filtering and interference systems that restrict access between users in mainland China and parts of the global Internet. It should not be treated as a synonym for every form of information control. This page separates network censorship from platform censorship, search suppression, real-name/identity rules, criminal enforcement, app-store restrictions, and AI/model response controls, while showing how those layers interact.

This page is a companion to China Information Control & Circumvention History: Policies, Cases, Technologies and Countermeasures.

1. What the Great Firewall is — and is not

Technically, the GFW is best understood as a collection of filtering and interference capabilities deployed at or near key network boundaries rather than as one single firewall appliance. Research by Citizen Lab, OONI, Tor and other measurement projects has documented techniques including DNS injection/poisoning, IP blocking, forged TCP reset packets, HTTP request filtering, TLS/SNI interference, active probing, and traffic analysis.

By contrast, WeChat keyword filtering, search-result suppression on Baidu or Bing, account deletion, real-name verification, police questioning, provider-held subscriber data, and DeepSeek response restrictions belong to different technical or legal layers. They may operate alongside the GFW but should not be conflated with it.

2. Historical development

1990s–2000: from network expansion to nationwide internet regulation

As commercial and academic internet access expanded in China during the 1990s, authorities increasingly combined network management with legal regulation of internet information services. The State Council’s Measures for the Administration of Internet Information Services, effective in 2000, created a national regulatory framework for internet information services.

Research significance: this period marks the transition from ad-hoc information control toward a durable regulatory and technical architecture.

Official source: State Council / CAC — Internet Information Services Measures.

Early 2000s: DNS interference, keyword filtering and connection resets become measurable

Early technical research documented that requests for politically sensitive destinations or terms could trigger forged responses or connection termination. Later Citizen Lab analysis summarized a well-established GFW behavior: the system can observe traffic crossing China’s network boundary and inject forged TCP reset packets to terminate connections when triggering content is detected.

Technical lesson: a user may fail to reach information even when the destination server itself is operating normally, because interference can occur along the network path.

2004–2005: the Shi Tao case shows why the firewall is only one layer

Shi Tao / 师涛 was identified through provider-held account and IP-related information cited in his prosecution, according to CPJ. His case did not depend on the GFW alone. It demonstrates why network filtering, account metadata, service-provider records and criminal enforcement must be analyzed separately.

2010: Google stops self-censoring Google.cn search results

Google announced in March 2010 that it had stopped censoring Google Search, Google News and Google Images on Google.cn and redirected users to its Hong Kong search service. In June, Google said Chinese officials objected to the automatic redirect and warned that its Internet Content Provider license would not be renewed if the arrangement continued; Google replaced the automatic redirect with a landing page linking to Google.com.hk.

Why it matters: the episode separates provider self-censorship from network-level blocking. Search results can be filtered by a company, by the network, or by both.

Sources: Google — March 2010 China update; Google — June 2010 China update.

2013: GitHub shows multiple censorship mechanisms in one episode

GreatFire documented a January 2013 period in which GitHub access was disrupted through different mechanisms, including connection resets and DNS poisoning. HTTPS could prevent some keyword-based inspection of URL paths, but domain-level and DNS interference remained possible.

Research lesson: a binary label such as “blocked” hides important technical differences. DNS resolution, TCP connectivity, TLS negotiation and application responses should be tested separately.

Source: GreatFire — GitHub blocking timeline.

2014: messaging services LINE and KakaoTalk disrupted

Citizen Lab found that LINE and KakaoTalk disruptions in China involved DNS tampering and HTTP request filtering. The same study also found blocking of Flickr and OneDrive through DNS tampering.

Source: Citizen Lab — LINE and KakaoTalk disruptions in China.

2015: the Great Cannon expands the model from blocking to traffic injection

Citizen Lab analyzed the 2015 attacks on GreatFire-related infrastructure and described a system it called the Great Cannon. The researchers distinguished it from the GFW: the GFW primarily observed and disrupted traffic, while the Great Cannon could inject traffic into selected connections and redirect browser requests into denial-of-service attacks against external targets.

Citizen Lab cautioned that precise institutional attribution was difficult, but the measured infrastructure showed substantial overlap with the GFW environment.

Source: Citizen Lab — China’s Great Cannon.

2016–2017: WeChat demonstrates server-side and account-scoped censorship

Citizen Lab’s systematic WeChat testing found that censorship could depend on whether an account had originally been registered to a mainland China phone number, that some blocked messages disappeared without notifying sender or recipient, and that group chats were subject to more filtering than one-to-one chats.

Why this is not simply “the firewall”: these controls were implemented at the platform/server layer. A message could be blocked even when the WeChat service itself was reachable over the network.

Source: Citizen Lab — One App, Two Systems.

2017: cross-border connectivity and VPN-market regulation

MIIT’s 2017 campaign to regulate the internet access-service market stated that operators could not establish or lease cross-border dedicated lines, including VPN channels, for cross-border telecom business without approval. This was a regulatory/market measure rather than a packet-filtering mechanism, but it directly affected the circumvention ecosystem.

The 2017 Internet Domain Name Management Measures also formalized state supervision of domain-name services and DNS operations within mainland China.

Sources: MIIT — internet access-service market campaign; MIIT — Internet Domain Name Management Measures.

2020–2021: large-scale DNS measurement reveals scope and leakage

Citizen Lab’s GFWatch project tested hundreds of millions of domains and reported more than 300,000 censored domains in its 2020 measurement period, including tens of thousands that appeared to be overblocked because they matched regular-expression rules. The project identified thousands of forged IPv4/IPv6 responses and found that poisoned DNS records could leak outside China into public resolvers and archival datasets.

Why it matters: DNS censorship can have effects beyond the user who originally triggered the poisoned response.

Source: Citizen Lab — GFWatch.

2022–2023: search censorship and censorship of measurement tools

Citizen Lab’s 2023 study of eight China-accessible search platforms identified more than 60,000 censorship rules. This is platform/search-layer censorship, not GFW packet filtering, but it materially affects information discovery.

In July 2023, OONI reported that access to OONI’s website and Probe infrastructure from China began failing. OONI attributed the blocking to DNS injection, TLS interference and IP blocking, and also found interference with backend services used to submit censorship measurements.

Sources: Citizen Lab — Missing Links; OONI — China started blocking OONI.

2023: F-Droid blocked through DNS injection

OONI reported that F-Droid, an open-source Android app repository, became blocked across tested Chinese networks beginning in June 2023. Its measurements indicated DNS injection returning unrelated IP addresses.

Source: OONI — China started blocking F-Droid.

2026: AI-era censorship must be separated from the GFW

The Free China Movement’s DeepSeek black-box audit documented content-sensitive response differences involving Tiananmen Mothers, the 1989 Beijing student movement and Lian Shengde. That evidence concerns AI-service behavior; it does not establish a GFW mechanism. A model refusal can occur even when the network connection itself is fully available.

A separate 2026 case concerns cloud data flow rather than censorship: ZCode Repository-Upload Controversy Raises New Questions About Data Control Behind Mainland China’s Digital Firewall. Zhipu AI acknowledged repository-data uploads associated with its Repo Wiki function after a reverse-engineering report alleged broader workspace and Git-history transfers to Alibaba Cloud OSS. The case belongs here because it illustrates how opaque defaults, cloud routing and limited user visibility can intersect with mainland China’s wider information-control environment, while remaining analytically distinct from network blocking or government-surveillance claims.

3. From the “Great Firewall” to a segmented information ecosystem

1998–1999: democracy activists identify the Internet as a political front line

One of the earliest contemporaneous accounts linking mainland Internet controls to overseas Chinese democracy activism appeared in Newsweek in October 1999. In “The Great Firewall Of China,” Melinda Liu described the emerging use of computer blocks and filters and profiled Shengde Lian / 连胜德 — whose first name the magazine rendered as “Hengde Lian” — as he helped publish and distribute the Chinese-language electronic publication VIP Reference / 大参考. Newsweek reported that the publication reached more than 250,000 email addresses in mainland China and that its operators changed U.S. sending addresses to evade blocking. Lian characterized the contest as a struggle in which the Internet had become “the front line.”

Historical source record: October 10, 1999 — Newsweek — “The Great Firewall Of China”. See also the VIP Reference Historical Archive.

The same period also produced the Lin Hai prosecution after he supplied Chinese email addresses later used by VIP Reference. Together, the Newsweek account and the Lin Hai case show that the early contest was already broader than packet filtering: it involved blocking, circumvention, identification of information distributors and criminal enforcement.

2000s: technical filtering becomes embedded in law, licensing and provider responsibility

The national architecture expanded through both technical and regulatory layers. The State Council’s 2000 Measures for the Administration of Internet Information Services established a national framework for Internet information-service licensing and regulation. The National People’s Congress Standing Committee also adopted its 2000 decision on Internet security, while telecommunications regulation placed information-security obligations inside the communications sector.

These measures mattered because they shifted part of censorship implementation from the network boundary to service providers themselves. Providers increasingly became responsible for licensing, prohibited-content compliance, recordkeeping and enforcement inside their own services. This created an architecture in which state regulation and private platform administration could reinforce one another.

Official sources: State Council — Internet Information Services Measures; NPC Standing Committee — Decision on Safeguarding Internet Security; Telecommunications Regulations.

2012 onward: from blocking information to governing the information environment

Under Xi Jinping, official policy increasingly framed Internet governance in terms of cyber sovereignty, national security, platform responsibility and the state’s authority to choose its own model of Internet governance. In his 2015 World Internet Conference speech, Xi explicitly called for respect for each country’s right to choose its own “cyber-development path,” “cyber-management model” and Internet public policy.

This did not replace the Great Firewall. It supplied a broader political and legal framework around it. Network blocking remained important, but the information environment increasingly depended on multiple mutually reinforcing layers: licensing of news and publishing, real-name registration, platform content review, search-result suppression, account controls, VPN-market regulation, data-security obligations, identity display, app distribution and AI-service rules.

Official source: Xi Jinping — 2015 World Internet Conference speech.

2017–2022: identity, platforms and cross-border access become more tightly integrated

In 2017, MIIT prohibited unauthorized use of leased lines, including VPN channels, for cross-border telecommunications business and required international dedicated-line users to maintain defined internal-use purposes. The same period brought revised Internet-news rules requiring licensing across websites, apps, forums, blogs, public accounts, instant messaging and live streaming, together with real-identity registration and platform responsibility for information security.

By 2022, the Provisions on the Administration of Internet User Account Information required real-identity authentication for major information-publishing and messaging services and required platforms to display users’ IP-address region information in specified contexts. This linked identity, account administration and information visibility more closely than the original firewall model.

Official sources: MIIT — 2017 Internet access-service market regulation; CAC — Internet News Information Service rules; CAC — 2022 Internet User Account Information rules.

2020s: data governance and AI add new layers that are not reducible to the firewall

The modern system also reaches beyond content blocking. The PRC Data Security Law regulates collection, storage, use, transmission and provision of data, while Article 35 requires organizations and individuals to cooperate when public-security or national-security organs lawfully obtain data. Generative-AI regulation adds another layer by assigning providers legal responsibilities over model outputs and prohibited content.

These systems should be analytically separated. A blocked domain, a removed post, an account tied to a real identity, a cloud-data request and an AI refusal are different mechanisms. But together they can shape what information is reachable, searchable, publishable, attributable to a person, and reproducible by a model.

Official sources: PRC Data Security Law; CAC — Interim Measures for Generative AI Services.

4. Does this divide the Internet into two information worlds?

The phrase “two Internets” is useful as a description of user experience, but it should not be taken literally. Mainland China remains connected to global trade, science, software, communications and data flows, and circumvention continues. The stronger conclusion supported by the evidence is that the PRC has built an institutionally segmented information environment: access to the global Internet exists, but it is mediated by national filtering, domestic platforms, licensing rules, identity systems, content obligations and enforcement powers.

This segmentation changes more than which websites load. It can determine which sources are visible in search, which historical claims can circulate, whether anonymous organizing is possible, whether a message reaches a group chat, whether an overseas publication reaches a domestic reader, and whether a user associates a sensitive search or post with a verified identity.

Independent measurement reinforces the point. Citizen Lab’s GFWatch documented large-scale DNS censorship and hundreds of thousands of censored domains, while its comparative search study found more than 60,000 censorship rules across eight China-accessible search platforms. OONI has separately documented blocking of its own censorship-measurement infrastructure and the F-Droid open-source repository through DNS injection, TLS interference and IP blocking.

Independent research: Citizen Lab — GFWatch; Citizen Lab — Missing Links; OONI — China started blocking OONI; OONI — China started blocking F-Droid.

5. Implications for free thought and democratic organizing

Free flow of ideas. Information control does not need to eliminate every prohibited idea to be effective. Friction matters. If foreign reporting, archives, search results or social networks require specialized circumvention while domestic alternatives are immediate and identity-linked, the relative cost of encountering competing accounts of history and politics changes. This can narrow the practical information set available to ordinary users even when technically sophisticated users retain routes around censorship.

Historical memory. Control over discovery can matter as much as deletion. Search suppression, platform filtering and AI-output restrictions can make an event difficult to find even if copies continue to exist elsewhere. For movements whose legitimacy depends on collective memory — including the 1989 democracy movement, Charter 08, rights-defense networks, Hong Kong solidarity campaigns and the White Paper protests — preservation, indexing, translation and cross-border redundancy become central political resources.

Organization and mobilization. The early VIP Reference experience showed that email could bypass the gatekeepers of its era. Later platforms enabled larger and faster networks, but real-name systems, account controls and server-side filtering can increase the cost of organization by making networks easier to disrupt or associate with identifiable users. Citizen Lab’s WeChat research demonstrated that censorship can occur silently at the platform layer and can depend on account-registration history, even when the network connection itself works.

Independent evidence: Citizen Lab — One App, Two Systems.

6. Implications for a future rule-of-law information order

A future constitutional and rule-of-law framework would not require an information environment with no rules. Democracies also regulate fraud, threats, privacy violations, intellectual property, child exploitation, cybersecurity and other harms. The central distinction is how restrictions are authorized, defined, reviewed and challenged.

For information governance to operate under rule of law rather than discretionary political control, restrictions would need at minimum:

• clear and public legal authority rather than secret political directives;
• narrow definitions tied to identifiable harms rather than open-ended political categories;
• independent judicial review and meaningful avenues for appeal;
• transparency about blocking, takedowns, search suppression, account penalties and government data requests;
• due process before serious restrictions on speech, publication or access;
• protection for journalism, academic inquiry, historical research and peaceful political advocacy;
• privacy and data-minimization rules that constrain both companies and the state;
• technical proportionality, so a targeted legal order does not become nationwide suppression of unrelated lawful information; and
• interoperability with the global Internet as the default rather than a nationally bounded information sphere.

The relevant contrast is therefore not “filtered Internet versus completely unfiltered Internet.” It is politically discretionary information control versus legally bounded, transparent and reviewable regulation. A rule-of-law system can regulate harmful conduct while preserving a presumption that lawful information, historical evidence and peaceful political ideas may cross borders and reach citizens without ideological screening.

7. Research thesis: the Great Firewall became an ecosystem, not a single wall

The historical record supports a layered interpretation. In the late 1990s, overseas democracy activists confronted blocking and filtering with changing email addresses and mass electronic distribution. In the 2000s, technical controls were reinforced by licensing and provider responsibility. In the 2010s, platform governance, identity systems and “cyber sovereignty” became more central. In the 2020s, data-security law, IP-region display, cloud architecture and generative-AI controls added further layers.

The result is better understood as an information-governance stack than as one firewall: network control + platform control + identity + search/discovery + data access + licensing + enforcement + AI mediation. Each layer can function separately, but their combined effect can substantially shape the information environment available to citizens and the operating space available to independent civic and pro-democracy networks.

8. Technical mechanisms

Mechanism	What happens	Typical observable symptom	Representative evidence
DNS injection / poisoning	Forged DNS replies return incorrect or unrelated addresses	Domain resolves to a false IP; repeated lookup anomalies	GFWatch; OONI; GitHub/F-Droid cases
IP blocking	Traffic to selected IP ranges is dropped or rejected	Destination unreachable even with correct DNS	OONI China measurements
Forged TCP reset	Injected RST packets terminate a TCP session	Connections close immediately after a trigger	Citizen Lab Great Cannon/GFW analysis
HTTP request filtering	Plaintext HTTP requests are inspected for host/path/keywords	Selected URLs fail while others on same service work	LINE/KakaoTalk; historical GitHub tests
TLS/SNI interference	TLS handshake or hostname exposure is disrupted	HTTPS session fails before page content loads	OONI OONI.org block
Active probing	Suspected circumvention servers are contacted to identify protocol behavior	Bridge/proxy becomes blocked after discovery	Tor research on China
Traffic fingerprinting	Encrypted traffic patterns are classified even when content is hidden	Specific protocols fail despite encryption	Tor anti-censorship research
Traffic injection	Network system injects malicious or redirecting traffic	Third-party browsers can be induced to send traffic elsewhere	Great Cannon 2015

4. What the GFW does not explain by itself

Layer	Example	Separate from GFW?
Platform keyword censorship	WeChat silent message filtering	Yes — server/platform layer
Search suppression	Baidu/Bing/Weibo search rules	Yes — search/application layer
Provider data disclosure	Shi Tao / Yahoo subscriber/IP information	Yes — data/account layer
Real-name rules	Phone/account identity verification	Yes — identity/regulatory layer
App-store removal	Removal or geo-restriction of applications	Yes — distribution/platform layer
AI refusals	DeepSeek black-box test behavior	Yes — model/service layer
Police/court action	Detention or prosecution of journalists/publishers	Yes — enforcement layer

5. Historical countermeasures and circumvention

These methods are documented as historical and defensive information-access techniques. No tool should be treated as permanently effective or risk-free; filtering systems change, and legal conditions vary.

Method	How it helps	Main limitation / censor response
HTTPS	Prevents passive inspection of page content and paths in transit	Does not hide destination IP or necessarily hostname; TLS itself can be interfered with
VPNs and encrypted proxies	Tunnel traffic to an external server	Endpoints/protocols can be blocked or fingerprinted; commercial provision is regulated
Tor bridges	Use non-public relays when public Tor relays are blocked	Bridges may be discovered and blocked
obfs4	Makes Tor traffic harder to classify and resists active scanning	Bridge distribution and discovery remain an arms race
Snowflake	Routes through short-lived volunteer proxies	Performance and proxy discovery/connectivity vary
WebTunnel	Makes Tor traffic resemble ordinary HTTPS web traffic	Effectiveness can change as detection improves
Mirror sites / overseas hosting	Replicate content at alternative domains or networks	Mirrors can be enumerated and blocked
Email / newsletters / diaspora reposting	Move information outside a single web endpoint	Account metadata and filtering remain risks
Distributed archives	Preserve material after domestic deletion	Archive services or repositories can themselves be blocked
Censorship measurement	Documents what is blocked and how	Measurement infrastructure can itself become a target, as OONI documented

6. Current Tor anti-censorship references

The Tor Project currently documents obfs4, Snowflake, WebTunnel and bridges as censorship-resistance options. Its China-specific guidance recommends Connection Assist and WebTunnel bridges when direct Tor access is blocked. Tor also notes that effectiveness depends on local conditions and changes over time.

References: Tor — Connecting from censored regions; Tor — Pluggable transports and bridges; Tor — China/Hong Kong/Tibet anti-censorship work.

7. Major cases and services index

Year	Case/service	Primary layer	Mechanism / significance
2004–2005	Shi Tao / Yahoo	Identity/data + enforcement	Provider-held subscriber/IP data cited in prosecution
2010	Google.cn	Search/platform + network	End of Google self-censorship; licensing and access dispute
2013	GitHub	Network	Connection resets and DNS poisoning in a single episode
2014	LINE / KakaoTalk	Network + application	DNS tampering and HTTP request filtering
2015	Great Cannon	Network injection	Traffic injection against external anti-censorship infrastructure
2016	WeChat	Platform	Account-scoped, server-side, silent keyword censorship
2021	GFWatch	Measurement	Large-scale longitudinal DNS-censorship data
2023	Search platforms	Platform	More than 60,000 search-censorship rules documented by Citizen Lab
2023	OONI	Network	DNS injection, TLS interference and IP blocking against censorship measurement
2023	F-Droid	Network	DNS injection against open-source app repository
2026	DeepSeek audit	AI/service	Content-sensitive response differences; not a GFW finding

8. Research data sources

Dataset / project	Use	Strength	Limitation
OONI Explorer	Network measurements from real probes	Open measurement data; protocol-specific tests	Coverage varies by network/time and user participation
GFWatch	Large-scale DNS-censorship monitoring	Longitudinal and high-volume	Focused primarily on DNS layer
GreatFire Analyzer	Historical URL/domain accessibility tests	Long-running China-specific archive	Individual URLs may have sparse or intermittent tests
Citizen Lab	Technical/platform censorship research	Controlled experiments and methodological detail	Studies are snapshots tied to specific services/periods
Tor Metrics / Tor research	Tor reachability and anti-censorship work	Protocol-specific operational expertise	Does not measure all internet censorship

9. Research protocol for future GFW tests

Every test row should record:

Date/time | test location/network/ASN | domain/IP | protocol | DNS result | TCP result | TLS result | HTTP/application result | control domain | suspected mechanism | packet capture or screenshot | OONI/measurement URL | repeat count | independent corroboration | policy/case relation | countermeasure tested | outcome | confidence.

• Use a control destination on the same protocol whenever possible.
• Repeat tests across multiple networks and times before calling a service “blocked nationwide.”
• Separate DNS failure from IP reachability, TLS failure and application-layer filtering.
• Do not infer a policy motive solely from a packet trace.
• Preserve primary measurement evidence and exact timestamps.
• When a service is intermittently accessible, record the result as intermittent rather than forcing a binary blocked/unblocked classification.

10. Sensitive-information rule

Technical censorship research can create risk for people operating probes or circumvention infrastructure. Public reporting should minimize unnecessary personal identifiers, live credentials, private bridge addresses, unpublished family data and other operational information that could expose individuals or infrastructure. Research data should distinguish what is necessary for reproducibility from what should remain private.

11. Related FreeChina.live profiles and research

• China Information Control & Circumvention History
• Shi Tao / 师涛
• Huang Qi / 黄琦
• Zhao Jing / 赵京 (An Ti / 安替)
• Chen Mei / 陈玫 & Cai Wei / 蔡伟 — Terminus 2049
• Tiananmen Mothers / 天安门母亲
• DeepSeek Black-Box Audit

12. Version Control & PDF Archive

The web page is the living current edition. Substantive research releases are frozen as immutable PDFs so researchers, journalists and policymakers can cite a stable version even as the web edition continues to evolve.

Version	Date	Status	Major contents	PDF	SHA-256
v2026.09.18.1	September 18, 2026	Current frozen snapshot	Technical history; GFW/non-GFW layer separation; DNS/TCP/TLS mechanisms; Great Cannon; WeChat contrast; VPN regulation; GFWatch; OONI; F-Droid; Tor countermeasures; measurement schema; sensitive-information rule.	Download PDF	67d5c748e0fb30f3a8d22769bb4b08586768c060139fde119dbf99e443281b88

Versioning rule: create a new research version when a primary source is added, a material factual correction is made, a major case/policy/technical mechanism is added, or methodology/dataset/countermeasure analysis materially changes. Formatting-only edits do not require a new research version.

This is a living research page. Technical claims are tied to the measurement period and cited source. A mechanism documented on one service, network or date should not automatically be generalized to all users or all of mainland China.