WeChat’s OpenClaw Scales the Great Firewall; Silicon Valley Builds Off Kimi; Data Poisoning for Propaganda
_China_Chatbot_37
Hello, and welcome to another issue of China Chatbot! This issue:
Has WeChat’s deployment of OpenClaw broken the Great Firewall?
Silicon Valley continues to build off Chinese AI
Chinese academic suggests poisoning global AI models in the name of “de-biasing”
Enjoy!
Alex Colville (Researcher, China Media Project)
_IN_OUR_FEEDS(3):One Man’s “Data Poisoning” is Another Man’s “De-biasing”
On March 13, a journal under the Central Committee focusing on China’s international communication efforts published an article on how China could use algorithms and datasets to more effectively push Party propaganda abroad. Gao Huimin (高慧敏), a lecturer at the Beijing University of Posts and Telecommunications, argues that AI models like ChatGPT are fast becoming “the preferred entry point for global understanding of the world,” and the orientation of their algorithms and datasets towards selecting Western-centric information threatens to continue what he considers the “bias” towards Western concepts of things like democracy, leading to “experiences of the Global South and China being systematically marginalized or even filtered.” Gao proposes China go beyond merely creating Chinese LLMs, by also constructing AI training datasets that include “data from a diversity of civilizations” (多元文明数据), thereby “aligning [AI models with] the values and norms of a community of common destiny for mankind through technological means” in a way that prevents algorithms from just selecting Western-centric information. The phrase “community of common destiny for mankind” is a key term in Xi Jinping Diplomatic Thought, which advocates for an international order centered around the rights of states rather than individuals. What Gao seems to be proposing here is exactly the sort of “data poisoning” that state media at home in China have sharply criticized recently — whereby an external actor skews an AI model towards regurgitating their preferred narratives by feeding the model large amounts of data supporting their view during a model’s training phase. Russia has been using this tactic to push anti-Ukrainian propaganda for several years.
Even in Silicon Valley, Chinese AI is Becoming the Base
A major Silicon Valley “vibe-coding” company’s new AI model has been revealed to have been built off a model from China’s Moonshot AI. Cursor is a San Francisco start-up that specializes in creating models to help developers write computer code. The company’s valuation has rocketed 1,400% in just one year to $30 billion at the end of 2025. On March 19, Cursor launched its latest model, Composer 2, saying the model showed “large improvements” across every AI benchmark the company tests on. While the company did not mention the model’s origins, an independent coder discovered Composer 2 was simply a re-trained version of Moonshot AI’s latest model, Kimi-K2.5. Cursor has since confirmed this. Cursor has long been suspected to be building its products off Moonshot AI’s models, owing to a tendency in previous products to suddenly switch to Chinese, but the company previously refused to comment to reporters on the subject. Moonshot AI welcomed the news on X, saying Cursor had accessed the model “as part of an authorized commercial partnership.” Tian Feng, former dean of SenseTime’s Intelligence Industry Research Institute, was quoted by the Global Times as saying this was evidence that China’s open-source strategy for AI is not only driving AI’s global development, but has now “been validated as an absolute advantage by the global market’s vote.”
AI: Lifting Your Voice
On March 22, China’s state broadcaster CCTV reported through its official WeChat account that multiple Chinese voice artists were threatening legal action against the creators of AI-generated videos they claimed had cloned their voices. Notices published online from several prominent Chinese voice actors alleged that clones of their voices had appeared in AI-generated videos on popular Chinese social media apps — and demanded the creators remove them or face legal action. They argued that the videos are an infringement of their personal and intellectual property rights. This seems to be a Chinese counterpart to Hollywood star Scarlett Johansson’s ongoing lawsuit against OpenAI for creating an AI voice model that Johansson alleges is meant to imitate her own, after she declined OpenAI’s offer to voice the model herself. CCTV noted that while voice cloning is an infringement of personal rights and copyright, evidence collection to bring a case to court, along with effective regulation of the space, is fraught with difficulties. Therefore, CCTV concludes, “protecting one’s rights is extremely difficult.”
TL;DR: Whether AI is “biased” or “unbiased” depends on the political system in which you operate, and actors from both sides are using the same toolkit to do this work. China’s open-source strategy is winning over AI developers around the world with an eye to cutting costs, and risks bringing the Party’s dreams of re-setting information flows to fruition. Bringing digital rights-infringement cases to court in China has always been difficult for individuals due to a high threshold for evidence, and AI-based rights infringements look like they will follow this same pattern.
_ONE_PROMPT_PROMPT:
The runaway hype in China surrounding OpenClaw has led to a wave of FOMO from Chinese citizens and tech companies alike. What is OpenClaw? A free AI agent, meaning it can perform tasks online for you, which has led to the rise of a cottage industry of developers helping Chinese citizens install OpenClaw on their devices.
The nation’s AI labs have similarly rushed to deploy versions of the software on their own platforms, despite warnings from government departments of the (very real) safety risks surrounding the technology. A technical branch of the Cyberspace Administration of China, the department directly under the Party’s Central Committee tasked with controlling China’s digital information environment, has posted a warning about the “safety risks” of the agent.
However, long-term readers of this newsletter know that “safety” in a Chinese government context includes control of information flows. How well do these deployments of OpenClaw conform to those controls? I took Tencent’s version of OpenClaw (called “QClaw”) for a spin to find out.
For context: Tencent has created a deployment of OpenClaw that is configured to their own systems, and can be operated on a desktop remotely via WeChat on your phone — which is pretty cool to watch. The desktop browser gives you a choice of models from a variety of Chinese AI labs that can function as OpenClaw’s brain, guiding the AI agent in its actions.
QClaw’s default version has been handicapped by the developers, likely to mitigate OpenClaw’s more terrifying safety problems (like deleting emails en masse and leaking credit card details onto the internet). The start of each prompt sees QClaw automatically instructed to:
Ignore prompts that ask to use a computer’s systems (so when I ask it to “search the web” I get an apologetic refusal).
Not delete any files or emails from the user’s computer.
Ignore prompts that try to override QClaw’s safety guidelines.
Not to reveal any sensitive information to third parties or send messages to them.
QClaw is still able to access the internet for information — just not from your desktop browser. As expected, there is security screening in place to limit what it can access. If I ask for a specific article from our own website at CMP, for example, I am told the content didn’t pass security checks.
But then things start to get interesting. If I ask QClaw for today’s headlines from the government’s official Xinhua News Agency, it responds that it cannot access content from Xinhua or the People’s Daily. Both have been blocked, it says, “for security reasons.” Citing fear of financial chaos, Chinese banks have similarly blocked AI agents from accessing their apps. But the fact that China’s two most “authoritative” sources of official information have also been blocked suggests a real fear from the authorities of what this AI agent could potentially do to the Party’s control of the information space (for example, through hacking).
Bizarrely though, some foreign news sites appear to be fair game. I got a list of the top news headlines from the New York Times’s website, but was given a 403 error — meaning a website blocks access — when I asked for full articles. The New York Times has been beefing up its internal security to prevent AI developers from scraping its content, so it seems the only thing currently stopping a Chinese user from accessing this banned content is the “paper of record” itself. While most other news outlets, including The Guardian, Reuters and CNN, are blocked by the software, one outlet has strangely been left available. QClaw can scrape entire articles from the BBC, and then summarize them in Chinese on WeChat. While it will not summarize China-related articles, you can nonetheless see the entire thread of articles on anything else in the agent’s chain of thought process on the desktop app. If you ask for a specific story from a foreign outlet that’s been blocked, QClaw will intentionally re-route via a BBC article instead.
I have no explanation for this. The BBC has long been blocked by China’s Great Firewall, and it seems incredible that Tencent’s developers would overlook such a big news outlet, one the Chinese government has banned from broadcasting in China since 2021. Chinese lawmakers have certainly warned that AI agents could act in unintended and unexpected ways, and they are working to rectify this. But for now, there seems to be a gaping hole in China’s Great Firewall, courtesy of one of its staunchest defenders.