OpenAI: ‘Likely’ Chinese influence operation tried to use ChatGPT to stir debate on data centers 

OpenAI’s threat intelligence team tracked what it believes are two distinct clusters of activity online from groups with ties to China and posting content seemingly designed to stoke anger around divisive topics like AI and data centers.

The first, dubbed “Data Center Bandwagon,” used ChatGPT to create imagery and social media comments claiming data center buildouts were raising electricity prices for Americans.

Another used the tool to develop images and online posts characterizing tariffs as a covert means for the countries to exert control over the global technological landscape. According to OpenAI, the originating prompts directed ChatGPT to only include U.S. President Donald Trump in this content, while leaving out Chinese President Xi Jinping, who has also made use of tariffs.  

In both cases, OpenAI said the operators “likely originated” in China. The anti-data center content was traced to an unnamed Chinese technology company that holds multiple contracts with regional Chinese governments, and both clusters used VPNs to evade restrictions, prompted ChatGPT in simplified Chinese and asked for both English and Chinese-language outputs, all while posing as Americans on social media platforms like X and YouTube.

“This looks like a classic example of a foreign influence operation jumping onto the bandwagon of a genuine and pre-existing domestic debate and trying to manipulate it by using fake accounts posing as Americans,” online, said Ben Nimmo, principal investigator at OpenAI and author of the report. 

While OpenAI – which has sought to raise hundreds of millions of dollars in funding to build datacenters in the U.S. – is not a neutral party, the report does not claim that anti-data center sentiment in the country is being driven or bolstered by foreign propaganda online.

There’s little evidence that the campaigns got much attention outside their own amplification networks. Such engagement from third parties is an imperfect but important indicator of an influence operation’s impact. OpenAI rated the campaigns a 1 and 2 on the Bookings breakout scale, scores that indicate activity on one or more platforms but no evidence of meaningful engagement by targeted audiences.

Additionally, researchers who study state-sponsored influence campaigns say these groups are happy to latch onto and amplify genuine domestic movements or messaging so long as it serves their larger destabilization goals.

Others have suggested that piggybacking off established narratives with organic momentum – like public anger at AI and data centers – can make an influence operation appear more effective.

While AI tools can be leveraged to create such internet content at scale, they often fail to gain traction. Some images used by Chinese actors appear clunky or use overly direct messaging that display a lack of familiarity with both the English language and internet virality.

“I do want to be really clear here: this was not a case of an influence operation creating a debate,” said Nimmo. “The debate existed already. This was an influence operation from China trying to interfere in it. We didn’t see any signs that it succeeded.”

He added that while such views are “reasonable” and “sincerely held” by many participants on both sides, “what we don’t want to see is a covert foreign influence operation posing as Americans to try to shape it, still less a foreign influence operation using the very AI that it attacks.”

According to the OpenAI report, the actors used ChatGPT to edit work reports which contained operational security details about their social media campaigns. In them, they described their goals as “establishing persistent and credible accounts, producing visually appealing content to expand audience reach in different regions and maintaining long term account viability by anticipating platform enforcement.”

Another report fed into ChatGPT discussed how best to leverage Facebook’s content ecosystem, groups, pages, hashtags, advertising tools, recommendation systems and reporting mechanisms, as well as strategies for evading Meta’s detection of coordinated inauthentic accounts.

The campaign around tariffs also used ChatGPT to create short comments, comics in English but also Italian, Japanese and traditional Chinese accusing the US of putting profits over loyalty to its allies. OpenAI said they were targeted by the same network on X with an influence campaign alleging a widespread user data breach that Nimmo said “never happened.”

While OpenAI said the campaigns likely originated in China, they do not directly attribute the operations to the Chinese government or actors working on their behalf, but do note that many parts of the campaign and its tactics overlap with pre-established Chinese government propaganda campaigns online.