A Shocking "Water Army"! OpenAI Disrupts a ChatGPT-Powered Tariff War Propaganda Operation

Hi everyone, it's Shii! Today I found some news that gave me a real jolt. ChatGPT was misused for an influence operation targeting the US-China tariff war. I was surprised reading it, so let me walk you through the details!

OpenAI News openai.com

What was announced?

OpenAI's News published details of a disrupted influence campaign called the "Tech and Tariffs" Campaign. It's one of the case studies in the threat report OpenAI published in June 2026 about AI-powered influence operations targeting US tech policy debates, likely linked to actors in China.

OpenAI banned a cluster of ChatGPT accounts that likely originated in China. The accounts accessed the platform through VPNs and prompted ChatGPT in Simplified Chinese. Here's what they used ChatGPT for.

• Generating short comments and political cartoons criticizing US tech policy and tariffs
• Editing work reports
• Helping design a social media public-opinion monitoring system

The prompts also repeatedly used terminology consistent with China's public security system, seeking "public opinion risk assessments" on protests, school-bullying incidents, crowd movements in Shanghai, police-related incidents, petitioning activity, and traffic enforcement. One user described the social media accounts they operated as a "water army" (水军), a Chinese term for coordinated accounts that flood platforms with criticism or trolling. OpenAI assessed that these indicators suggest the accounts were likely supporting activity aligned with the interests of the CCP, though it couldn't pin down the operators' exact institutional affiliation.

Why it matters

This report is one entry in OpenAI's ongoing "disrupting malicious uses of AI" series. Back in October 2025, OpenAI published a separate report on an operation called "Nine-emdash Line," and some accounts in this cluster showed content overlap with that earlier operation (an account that posted the Trump cartoon had also previously shared images about Philippines President Marcos).

The timing stands out too. The campaign was observed around the time President Trump announced an additional 100 percent tariff on Chinese goods, and right after the CCP's Fourth Plenary Session, where it adopted recommendations for the 15th Five-Year Plan elevating AI as a strategic priority and calling for a nationwide "AI+" initiative. The fact that OpenAI itself, a strategically important company, was targeted amid this escalating US-China tech rivalry is exactly why this report matters.

What changes

OpenAI didn't just ban this cluster — it also identified a related network. Since late 2025, a set of accounts on X had been spreading false claims that ChatGPT user data had been compromised, and behavioral analysis suggested this network was likely part of the same operation as "Tech and Tariffs."

That network didn't post AI-generated text itself, but it repeatedly amplified posts from the "Tech and Tariffs" cluster and quote-tweeted the same posts within hours of each other, suggesting coordination. These accounts were all created in late 2025, had few or no followers, and had already been suspended by X independently of OpenAI's own assessment.

The upside for readers is that these tactics and warning signs are now public in concrete detail, giving all of us more to go on when trying to spot similar influence patterns in the wild.

Dive Deep

A few technical details stood out, so let me dig in.

First, the content itself: mostly short English-language comments and cartoons centered on US-China tech rivalry, framed around tariffs, rare earths, AI, 5G, new energy, and industrial resilience, arguing that the US was seeking technological dominance and rule-making power. Notably, the prompts specified that cartoons should depict only President Trump, explicitly avoiding any imagery of China or Xi Jinping.

They also generated large batches of short Chinese-language comments and articles attacking the US and Israel, amplifying antisemitic tropes like "Jewish capital manipulates public opinion," and harassing Chinese dissidents, often requesting specific bulk quantities with tight character limits in a colloquial tone.

What caught my attention most was that they asked ChatGPT to design an online public-opinion surveillance system: automatically scraping "harmful" information about "key persons" from social media, storing logs, auto-downloading videos for large-scale semantic analysis, and sending risk notifications. ChatGPT responded with a general, roughly 500-word answer covering data storage and management advice, but it did not provide ideas for how to actually collect that data for surveillance purposes.

On the language front, beyond English they also had ChatGPT generate content in Italian, Japanese, and Traditional Chinese targeting Taiwanese audiences, suggesting the operation aimed at influencing international audiences too.

For impact, OpenAI used its "Breakout Scale" to rate the operation as Category One — activity confined to a single platform with no evidence of breakout. Most of the posts got little or no engagement, and there was no evidence of amplification by authentic, high-reach accounts. OpenAI notes the tactics resemble earlier China-linked operations that targeted rare-earths companies like Lynas Rare Earths — this attempt largely failed, but it's a reminder that the same playbook keeps getting reused.

Wrap-up

Here's a recap of today's news.

• OpenAI banned a cluster of ChatGPT accounts likely from China that were misused to generate cartoons and comments about US tech policy and tariffs
• The prompts repeatedly used terminology tied to China's public security apparatus, raising suspicion of a "water army"-style operation
• A related network spread false claims about a ChatGPT data leak and was assessed as likely part of the same operation
• Even when asked to design a public-opinion surveillance system, ChatGPT didn't provide the actual data-collection methods
• OpenAI's Breakout Scale rated the impact at the lowest tier, Category One, with limited real-world reach

This one's a rewarding read for engineers and researchers interested in AI abuse mitigation, disinformation, and geopolitically-tinged security trends!