OpenAI bans Chinese 'water army' using ChatGPT to push anti-US tariff cartoons

OpenAI has pulled the curtain back on a coordinated influence operation it’s calling “Tech and Tariffs,” banning a cluster of ChatGPT accounts that likely originated in China. The accounts weren’t trying to build the next killer app—they were running a digital “water army” (水军), a term one user explicitly used, to flood social media with AI-generated political cartoons and short comments criticizing US tech policy, tariffs, and trade restrictions. The operation was detailed in OpenAI’s June 2026 threat report, and it paints a picture of a pragmatic but unsophisticated attempt to tilt the global conversation on tech competition.

The operators used VPNs and prompted the models in Simplified Chinese, generating content in English, Italian, Japanese, and Traditional Chinese to target a global audience. Their thematic focus was sharp: frame the US as a backstabbing hegemon obsessed with rule-making power and profit over loyalty. Their prompts specifically instructed the AI to depict President Trump in cartoons, while explicitly forbidding any image of China or Xi Jinping. Beyond the visual propaganda, they churned out batches of anti-US, anti-Israel content and amplified ugly anti-Jewish tropes, such as “Jewish capital manipulates public opinion,” alongside harassment of Chinese dissidents.

The behavior wasn’t limited to content creation. The accounts also probed ChatGPT for advice on designing an AI system to automatically surveil “harmful” information from “key persons” on social media, though OpenAI notes the model only returned generic data storage advice, not a surveillance blueprint. The linguistic fingerprints were telling: the accounts consistently used terminology linked to China’s public security system, requesting risk assessments on protests, crowd movements in Shanghai, and police incidents. Yet, OpenAI stops short of a definitive institutional attribution, saying they “could not establish the operators’ precise institutional affiliation,” even if the activity clearly aligns with CCP interests.

A separate, intertwined network on X made false claims that ChatGPT user data had been compromised. This smear campaign against OpenAI didn’t use model-generated text, but its accounts actively amplified and reposted content from the cartoon-posting “Tech and Tariffs” cluster, even quote-tweeting the same unrelated verified account within hours. It’s a classic muddy-the-waters tactic. The whole operation reveals a stark reality of modern information warfare: you don’t need a zero-day exploit when an off-the-shelf generative AI tool and a batch of sock puppet accounts can seek to erode trust in a commercial product and a rival nation’s policy simultaneously.