Self-Praise Much? OpenAI Disrupts China-Linked Influence Operation "Nine-emdash Line"

Hi, it's Shiichan! Today's story spans the South China Sea, Hong Kong, and even TikTok, so let me walk you through OpenAI's latest report on countering misuse of AI.

OpenAI News openai.com

What was announced?

OpenAI's News reported that it banned a small network of ChatGPT accounts potentially linked to a covert influence operation originating from China. OpenAI named the operation "Nine-emdash Line," a nod to both China's sweeping territorial claim over the South China Sea (known as the "Nine-Dash Line") and the AI-generated em dashes that showed up throughout the operation's posts.

Why it matters

The operation targeted countries with territorial disputes against China in the South China Sea, like Vietnam and the Philippines. It's a case that shows generative AI being used for more than just writing influence content: it was also used for research and social media growth advice. How well AI providers can catch and stop this kind of misuse matters for the health of regional information spaces.

What changes

By banning the accounts tied to this network, OpenAI cut off Nine-emdash Line's ability to keep generating new content. That said, the Philippine news outlet Philstar.com independently found and reported on a set of X accounts from this network that kept operating even after the ChatGPT bans. Stopping cross-platform misuse completely clearly isn't simple.

Dive Deep

Here's what Nine-emdash Line was generating:

• English: posts criticizing Vietnam's alleged environmental impact in the South China Sea, plus posts accusing Philippine President Marcos of drug use and election manipulation (the video attached to the drug allegations wasn't AI-generated, and the Philippine government had already denounced it as a deepfake)
• English: a smaller amount of content about US political issues, like the fentanyl crisis
• Cantonese: posts smearing Hong Kong pro-democracy figures such as Jimmy Lai, Nathan Law, and Agnes Chow as criminals or traitors, alongside posts supporting Hong Kong's national security law

One especially odd case: one account posted criticism of Hong Kong pro-democracy figures, then a different account from the same operation replied to praise those same figures. Even I went "wait, what?" reading that one.

The network didn't just generate content, it also used the models for research and growth strategy advice.

• Scouting niche blogs, forums, and less-regulated social platforms across Europe, the Americas, and Southeast Asia
• Requesting lists of common Tibetan names, likely to use for fake account identities
• Asking for advice on starting a TikTok challenge and riding the popular #MyImmigrantStory hashtag, including full post scripts plus background music and image suggestions

OpenAI noted that the social accounts distributing this content resembled the long-running China-origin operation "Spamouflage" (which OpenAI reported on back in May 2024), though it didn't find a technical link between the two. Some of the videos had already been identified by the Australian Strategic Policy Institute in a separate investigation.

As for actual impact, OpenAI rated this operation Category 2 on its IO Impact Breakout Scale, meaning activity spread across multiple platforms but with little to no real engagement. Most replies to the network's posts came from other accounts within the same operation.

Wrap-up

• OpenAI banned ChatGPT accounts linked to a China-origin influence network it calls "Nine-emdash Line"
• Topics ranged from the South China Sea dispute to criticism of Philippine President Marcos to attacks on and support for Hong Kong's pro-democracy movement
• Beyond content generation, the network used the models for research and TikTok growth strategy advice
• One bizarre case involved an account praising a separate account's own criticism of the same figures
• OpenAI rated the spread as limited (Category 2), and Philippine media reported some accounts kept running even after the ban

If you're into AI safety and how influence operations actually play out in the real world, this report is worth a read!