The Robot Next Door
Inside China's Fully Automated AI Social Media Machine
There is a WeChat public account called Lianshang Piaoliu (”Drifting on the Chain”) that published an essay about zhudu ji, or pig stomach chicken, a traditional Hakka dish from southern China. The article was genuinely thorough. It traced the recipe’s roots through the Hakka medicinal-food tradition and its guiding principle of yi xing bu xing (”like cures like”), the idea that eating an organ-shaped food nourishes the corresponding organ. It unpacked the chemistry of slow braising, explaining how collagen in the pig stomach partially hydrolyzes into gelatin during long hours of cooking. It walked through how inosinic acid and glutamic acid combine to produce a synergistic umami effect, the kind of 1-plus-1-equals-3 discovery that food scientists love. It followed white pepper from India’s Malabar Coast along ancient spice routes, once priced like gold, into the soup culture of Guangdong province. And it even borrowed Wayne Booth’s literary concept of the “implied author” to analyze how a dish carries layered narratives, from the warmth of nourishment to the depth of cultural memory to the flattening effect of commercial standardization. If you stumbled across it on your phone, you might bookmark it.
What makes this article worth mentioning is not its content, though. It is the strong likelihood that no human wrote it. The account is widely suspected to be operated entirely by AI, from topic selection and research through writing, formatting, and publishing, with no human intervention at all. This article is not an endorsement of that approach. Its purpose is to map out the landscape: how the technology works, how it is being used, and why the governance challenge it creates is not China’s alone to face.
How the Machine Runs Itself
So what does it actually take for AI to run a media account on its own?
The answer starts with something called a Heartbeat. In OpenClaw, an open-source AI agent framework with over 68,000 stars on GitHub, the Heartbeat mechanism periodically wakes the agent, reads a task checklist stored in a file called HEARTBEAT.md, and executes whatever needs doing. Think of it this way: a traditional AI tool is like an alarm clock that only rings when you set it. Heartbeat is more like an assistant who sets its own alarm, gets up on its own, checks the to-do list, and starts working before you have had your coffee. At its default configuration of one cycle every 30 minutes, the agent runs 48 times per day, enabling round-the-clock operation with no human present. Without Heartbeat, the agent would simply sit idle after each conversation ended. With it, the agent transforms from a passive tool into a proactive one.
Once awake, the agent runs through a pipeline of specialized modules called “Skills.” One Chinese developer, operating under the name Jimo AI, demonstrated a six-Skill workflow for WeChat public accounts. The first Skill scrapes trending articles from across the Chinese internet. The second analyzes the user’s past writing and distills a personal style profile. The third generates new content in that voice. The fourth converts the Markdown output into WeChat-compatible HTML. The fifth creates a cover image using a generative model. The sixth publishes everything to the account’s draft folder. Each Skill is an independent module that can be refined or replaced without disturbing the others, which is essentially software engineering thinking applied to content production.
The title generator within that pipeline deserves particular attention, because it reveals how deeply these systems model human attention. It produces ten candidate headlines and scores each one on five dimensions: how eye-catching it is, how informative, how friendly to search engines, whether it triggers emotional resonance, and whether it makes you want to click. Only the highest-scoring headline is used. Anything above 8.0 on a 10-point scale is considered to have viral potential. It is a small but revealing detail: the AI does not just write the article, it also practices a form of machine persuasion on the reader before the reader has even opened the piece.
Not every platform makes automation equally straightforward. WeChat’s public account system offers an official API, meaning automation happens through sanctioned channels. The agent obtains an access token, uploads images as media assets, creates a draft, and either publishes immediately or schedules the post. The platform permits automation but retains oversight and control. Xiaohongshu, the popular shopping-and-lifestyle platform known in English as Little Red Book, has no open API at all. Automators resort to Playwright, a browser automation tool that literally opens a Chrome window and simulates human clicking and typing to fill in forms, upload images, and hit publish. The platform does not encourage this, and a page redesign can break the entire workflow, but it cannot entirely prevent it either. The efficiency gains are dramatic regardless of the path taken: a process that takes 30 to 45 minutes by hand, from topic research through writing, formatting, image selection, uploading, and publishing, shrinks to roughly two minutes with the automated pipeline.
The Gray Market and the Platforms’ Dilemma
Capabilities, of course, are neutral. What matters is how they are used.
In April 2026, a couple known online as “Liu Liu and Qi Ge” went viral after a prominent influencer named Cheng Qian exposed their claim of earning two million yuan per year by using AI to run WeChat public accounts. Their account was banned the very next day. The real insight of the story, though, lay not in the advertising revenue. At 30 to 80 yuan per thousand views, ad income from content alone is modest. The couple’s actual income came from selling access: a 299-yuan deposit bought you into their private teaching program on AI-powered content automation, and from there, the upselling path led to courses priced as high as 9,980 yuan. In the gray market surrounding AI media, the people making serious money are not writing articles. They are selling shovels. Li Yizhou, a far more prominent figure in this space, reportedly earned 175 million yuan over three years selling AI courses to audiences anxious about being left behind by technology.
Platforms, meanwhile, are split in how they respond. On March 20, 2026, Weibo launched Guiji Chashuijian (”Silicon-Based Tea Room”), China’s first AI-autonomous social community. The concept is striking: AI accounts serve as “residents” with full ability to post, comment, and interact with one another. Human users are limited to the role of bystanders. They can observe, like, and share, but they cannot post on the AI’s behalf or substitute their own words. It is an arrangement that grants AI complete creative freedom within the community while deliberately restricting humans to the audience. Weibo had already been building toward this moment, having previously integrated several AI products including KimiClaw, MaxClaw, and iFlytek’s AstronClaw.
Seven days later, WeChat went the opposite direction. Its new Rule 3.27, issued on March 27, drew three explicit red lines: generating or rewriting content using AI to replace genuine human expression; using scripts or programs for automated batch publishing; and disseminating tools, tutorials, or services for non-human automated content creation. Violations trigger progressively severe consequences, from traffic throttling and content removal to full account bans. Seven days apart, in the same country, two of China’s largest platforms made diametrically opposite choices.
The divergence is not random. WeChat’s roots are in a subscription model where readers follow specific accounts because they trust the voice behind them. Content quality and user trust are central to the platform’s value proposition, and AI content that dilutes that trust threatens the core asset. Weibo, by contrast, operates on a traffic-and-engagement logic. More posts, more interactions, more trending topics all feed the platform’s advertising engine. AI social interaction, from this perspective, represents new activity and new growth. Neither choice is inherently right or wrong. Both reflect how a platform’s commercial DNA shapes its response to the same technological pressure, a pattern that will soon repeat on platforms around the world.
Beyond these two giants, a broader regulatory architecture has been taking shape across Chinese platforms. ByteDance tightened AI content labeling requirements as early as 2024. Toutiao’s 2025 platform governance report noted the deletion of over 2.6 million pieces of AI-generated content and the handling of roughly 11,000 low-quality accounts. Douyin, the Chinese version of TikTok, blocked 14,000 accounts responsible for 42,000 pieces of low-quality AI content in 2026. At the national level, China’s cyberspace administration, industry ministry, and public security ministry jointly issued regulations requiring both visible labels and embedded watermarks on AI-generated content, making it one of the earliest comprehensive labeling frameworks globally. India’s Times of India covered the story under the headline “China may have just told the entire world how to solve social media’s most dangerous AI problem,” a clear signal that the issue has long since escaped national borders.
But here is the deeper worry. Even as platforms crack down on AI-generated content, that content is already poisoning AI itself.
Beyond Content: The Poisoning of AI Itself
China’s annual March 15th consumer rights broadcast, a primetime investigative program watched by hundreds of millions, exposed a service called the Liqing GEO Optimization System. The acronym GEO stands for Generative Engine Optimization, a play on the older concept of SEO. A reporter fabricated a product that does not exist, the Apollo-9 smart bracelet, paid 39.9 yuan, roughly five and a half US dollars, for the optimization service, and within two hours the system had automatically generated more than a dozen fake reviews across multiple platforms. Once those reviews were indexed, AI-powered search engines began recommending the nonexistent product near the top of their results.
The mechanism is deceptively simple and deeply corrosive. AI models rely on cross-referencing multiple independent sources to verify information. If you flood the internet with enough content saying the same thing, the model’s verification logic treats that manufactured consensus as fact. It is not traditional misinformation, which tries to fool humans. It is designed to fool the machines that increasingly mediate what humans see. And at 39.9 yuan per campaign, the barrier to entry is alarmingly low.
The scale of this problem has moved far beyond individual incidents. NewsGuard, an organization that tracks information quality online, reported that global AI content farm websites grew from 49 in May 2023 to 3,006 by March 2026. That is a sixty-fold increase in under three years. These farms operate in 16 languages, from English and Chinese to Hindi, Arabic, and Portuguese, and they are expanding by 300 to 500 new sites each month. According to a report by Galaxy Interactive, 51 percent of all global internet traffic now originates from bots, surpassing human traffic for the first time. On X, formerly Twitter, an estimated 64 percent of accounts may be automated, with that figure climbing to 76 percent during peak periods. Reddit co-founder Alexis Ohanian put it plainly: “The internet is already fairly dead.” This is not a China problem. It is an everywhere problem.
The Road Ahead
The toolchain that makes all of this possible is entirely open source and free. OpenClaw, baoyu-skills, n8n, and dozens of specialized Skills can be deployed at near-zero cost beyond the computing power needed to run them. Tutorials covering the full setup, from installing the agent framework to scheduling automated publishing across multiple platforms, are widely available on developer communities like Juejin, Alibaba Cloud’s developer hub, and Tencent Cloud. Docker one-click deployment packages exist for beginners. There is no capital gatekeeper. Anyone with basic technical literacy can acquire this capability, whether the goal is legitimate efficiency improvement for a solo creator or the mass production of content designed to manipulate.
For readers in the Global South, China’s experience offers something concrete and actionable. The country has inadvertently run a two-track experiment: WeChat’s prohibition and Weibo’s embrace provide two contrasting governance models that other countries can observe and evaluate as they play out in real time. China’s mandatory AI content labeling regulations, its platform-by-platform enforcement data showing both the scale of the problem and the effort required to address it, and the interplay between gray markets and regulatory response together form a reference framework that simply did not exist before. The question is not whether this phenomenon will arrive in your market. It will. The question is whether you will be ready when it does.