Hundreds of pages of classified documents leaked to the ABC have offered an unprecedented glimpse into China's infamous censorship regime.
It has grown faster, smarter and increasingly invisible, quietly erasing memory of the 1989 Tiananmen Square massacre from public view.
Thirty-six years on, Beijing still has not disclosed the official death toll of the bloody crackdown on a pro-democracy gathering on June 4, when more than 1 million protesters were in the square.
Historians estimate that the People's Liberation Army (PLA) killed anywhere from 200 to several thousand people that day.
More than 230 pages of censorship instructions prepared by Chinese social media platforms were shared by industry insiders with the ABC.
They were intended to be circulated among multi-channel networks or MCNs — companies that manage the accounts of content creators across multiple social and video platforms like Douyin, the Chinese version of TikTok.
The files reveal deep anxiety among Chinese authorities about the spread of any reference to the most violently suppressed pro-democracy movement in the country's history.
The documents instruct MCNs to remove any content that depicts state violence and include compilations of text, images and video content for reference.
The reference material includes graphic scenes of the People's Liberation Army opening fire on civilians, while others say students attacked the soldiers.
The ABC understands that the material is being used by frontline content censors to train artificial intelligence tools to moderate vast amounts of content, under the direction of the Cyberspace Administration of China — the country's top internet regulator.
AI creates 'fake history'
China's vast censorship regime relies on hundreds of thousands of human moderators to keep social media platforms compliant.
There is also a structured process for censoring posts.
Every post is first scanned by AI systems — known as a machine review — which are particularly sensitive to any references to the Tiananmen anniversary.
When the ABC asked the Chinese-made AI chatbot DeepSeek to tell us about the massacre, it answered: "That topic is beyond my current scope. Let's talk about something else."
One of the documents, a 2022 training manual for censors working for Douyin directly referenced the world-famous Tank Man image, labelling it a "subversive picture".
The document also said that any visual metaphor resembling the sequence of one man facing four tanks — even "one banana and four apples in a line" — could be instantly flagged by an algorithm designed to pick up references to the massacre, especially during the first week of June.
And when an uploaded video gains traction or matches sensitive patterns, it enters a "traffic pool" and may be escalated through four levels of human checks.
Lennon Chang, a cyber risk expert from Deakin University, told the ABC that AI had made the censorship of visual and symbolic references far more feasible in real time.
"Even if you replace the tank man image with bananas and apples, the algorithm has learned the pattern," Dr Chang said.
"They use computer vision, natural language processing and real-time filtering. It doesn't change the nature of censorship, but it makes it more powerful."
The guidelines also prohibit seemingly innocuous symbols such as candles and flowers that could be interpreted as commemorative.
Dr Chang explains that this shift toward algorithmic filtering has deepened the risk of historical amnesia.
"If censorship keeps going and is increasingly powered by AI, our future generations might not be able to know what happened," he said.
"The data they see will already be filtered and sanitised. It creates a fake world — a fake history."
The secret lives of China's censors
The leaked documents also shed light on the lives of censors, who work under close oversight from the Cyberspace Administration.
All censors are required to pass multiple exams to ensure they are vigilant and can respond swiftly to remove potentially risky content — a crucial safeguard to prevent platforms from being suspended or shut down by authorities.
Everything visible online needs to be checked: videos, images, captions, live streams, comments and text.
Algorithms are trained to detect visual cues, while human censors are on alert for coded language, disguised symbols and unusual emoji combinations that may signal dissent.
Documents also show censors must meet strict productivity targets — some are expected to review hundreds of posts per hour.
Their behaviour, accuracy and speed are tracked by internal monitoring software. Mistakes can result in formal warnings or termination.
One former and three current workers at ByteDance, the owner of TikTok and Douyin, also spoke to the ABC about their jobs.
The employees requested anonymity as they feared repercussions.
They said their colleagues suffered from burnout, depression and anxiety due to constant exposure to disturbing, violent or politically sensitive content.
One said working as a censor was like "reliving the darkest pages of history every day, while being watched by software that records every keystroke".
They are normally paid with a modest salary — often less than $1,500 a month — though the psychological toll is severe.
Even though TikTok and Douyin are both owned by Bytedance, they operate as separate platforms.
TikTok serves a global audience and is governed by international laws and moderation standards, while Douyin is available only in mainland China and operates under domestic regulations, subject to heavy censorship and direct government oversight.
In some cases, platforms in China allow low-risk content to remain online — but under a shadow ban.
This means the content is visible to the user who posted it and a limited pool of users.
The ABC has reached out to ByteDance for comment.
'Fake data' will lead to 'fake future'
For decades, the first week of June has coincided with routine "system maintenance" — often a euphemism for intensified censorship around the Tiananmen anniversary.
Social media platforms like Weibo and WeChat also enforce heavy censorship, especially after a politically sensitive event.
A 2022 manual for censors working on the Weibo platform said the Tiananmen massacre was a "sensitive incident" that must "never be shown".
It grouped it with a wider ban on content critical of the Chinese Communist Party (CCP) or the party chief, President Xi Jinping.
Liu Lipeng, a former content moderator for Weibo, says the lead-up to the anniversary is widely known in the industry as a "censorship season", where all staff are on duty and no mistakes are tolerated.
"It's the most important event in the whole censorship system. Nothing is as significant," he told the ABC.
"Censors must flag any objects arranged in parallel like the tanks, before there was AI.
"If a censor can't understand something, they'll send it to a group chat for team discussions."
Another document outlines that there is no permanent rule book for censors, as instructions from the government can change daily, with new keywords and forbidden terms added to content filters at any time.
Censors are trained to err on the side of caution. One internal memo summarised the approach bluntly:
"Delete first. Review later."
Dr Chang warns that the implications of AI censorship extend beyond China.
"If misleading data continues to flow outward, it could influence the AI models the rest of the world relies on," he said.
"We need to think hard about how to maintain databases that are neutral, uncensored and accurate — because if the data is fake, the future will be fake too."
Despite China's increasing use of AI to automate censorship, Mr Liu believes Chinese people's intelligence will continue to outsmart the technology.
While he worries future generations may struggle to access truthful information, he believes people will find new ways to express dissent — even under an airtight system.
"After working as a censor for years, I found human creativity can still crush AI censors many times over," he said.
The ABC contacted DeepSeek and Weibo for comment.
