China Used ChatGPT to Fake a U.S. Power-Bill Revolt

OpenAI banned two China-origin ChatGPT clusters in June 2026 for writing fake American posts about data-center power bills and U.S. tariffs. The operators prompted in Simplified Chinese, hopped VPNs, and asked the model to sound like U.S. mothers, clerks, students, and investors. Ben Nimmo, OpenAI’s principal investigator, told reporters it was a twist that they used American AI to do it.

The ads did not travel. The bill fight they copied was already underway, and it did not need a chatbot in Beijing to exist. That is the part that lasted.

OpenAI Banned Two China-Linked ChatGPT Clusters

On June 10, 2026, OpenAI said it had banned two clusters of ChatGPT accounts that likely originated in China after they used the model in support of covert influence work aimed at a live American argument over AI policy and tech rules. The company named the first cluster the Data Center Bandwagon campaign. It named the second Tech and Tariffs.

OpenAI does not allow access to its models from China, so the operators used VPNs. They asked for English and Chinese output while writing the prompts in Simplified Chinese. The posts went up on X through accounts OpenAI judged likely inauthentic, sometimes next to links to real news about grid auctions and data-center power demand.

THE RUN OF THE BANS

1. October 2025: President Donald Trump announces an additional 100 percent tariff on Chinese goods, and the Chinese Communist Party’s Fourth Plenum elevates AI in the next five-year plan.
2. Late 2025: Inauthentic X accounts appear. ChatGPT prompting for English comments, comics, and cartoons is underway.
3. Early 2026: Data Center Bandwagon posts land on X with hashtags such as #capacityauction and #datacenters.
4. June 10, 2026: OpenAI publishes the findings, bans the clusters, and Rep. John Moolenaar and Sen. Tom Cotton answer the same day.

OpenAI stopped short of pinning either cluster on the Chinese government as a named command. The Data Center Bandwagon operators, it said, were likely a social media team at a private Chinese technology company doing work for provincial-level government clients, in a commercial trade that supports Party-state public-opinion work. For Tech and Tariffs, prompts used language associated with China’s public security system, and one user called the accounts a water army. OpenAI said it still could not fix a precise institutional home.

Fake Americans, Real Utility Clippings

The Bandwagon cluster did not invent a grievance. It asked ChatGPT for comic strips about capacity auction prices drawn from a legitimate regional paper, then framed peak demand as a data-center problem whose cost landed on ordinary households. Operators also had the model add text to generic electricity-market images so the pictures said regular people were subsidizing AI plants.

They wanted the voices to look local. OpenAI said the same cluster posed as U.S.-based Chinese immigrants and as workers, students, mothers, clerks, and investors. A second track went after overseas Chinese audiences, including requests for insults aimed at the dissident Li Ying, known as Teacher Li, and at other Chinese political commentators. In those cases the model refused to generate inflammatory personal attacks.

The same accounts treated ChatGPT as a back-office clerk. They asked for code to automate logins and to manage several social platforms at once, and they used it to strip usernames, prepend X or YouTube links, and format worksheets. One internal-style report they uploaded described Facebook operations in detail, including lifestyle personas, backup accounts, and a dual track of organic posts plus ads, all aimed at staying up long enough to look real.

WHAT THE TWO CLUSTERS PRODUCED

Nimmo told reporters this was not an operation that created a debate. It jumped onto one that already existed, then tried to steer it with fake Americans. OpenAI had seen other China-linked misuse of its models. It had not, he said, seen those models used this way on the U.S. data-center fight.

Draw Trump, Never Xi

The Tech and Tariffs cluster had a stricter art brief. Operators asked for short comments and cartoons that framed U.S. tariffs, rare earths, AI, 5G, new energy, and industrial policy as a bid for dominance and rule-making power. They specified that the cartoons should depict only President Trump and should not include any imagery of China or of Xi Jinping.

OpenAI published samples that match that brief, including Trump swinging a hammer at a wall labeled Global Future and sawing a ladder while standing on it. The same cluster generated large batches of short comments in Chinese that favored Beijing, attacked the United States and Israel, and recycled anti-Jewish tropes. It also asked for bulk comments in Italian, Japanese, and Traditional Chinese, the last aimed at Taiwan.

The targeting of OpenAI and US data center buildouts is significant not because the operation appears to have shifted public opinion, but because it shows PRC-origin influence operators testing narratives against AI infrastructure.

OpenAI, June 2026 threat report

A related set of likely inauthentic X accounts spent late 2025 claiming ChatGPT user data had been stolen and that the theft had wrecked their lives. OpenAI said those allegations were entirely false. The company assessed, from open-source behavior, that the data-theft accounts were likely part of the same X network as Tech and Tariffs. None of the theft posts appeared to have been written by its models. X later suspended the accounts on its own.

One X account that posted a ChatGPT-made Trump cartoon had also shared images of Philippines President Ferdinand Marcos Jr. that OpenAI had previously tied to an operation it called Nine-emdash Line in October 2025. That overlap, the company said, was not enough to prove the two operations were the same shop.

The Power Bill Fight Was Already Real

The reason the comics could be written at all is that U.S. households were already arguing about rates. OpenAI said the operators tried to exploit existing public concern about energy prices and the local impact of data-center construction. Congressional researchers have been mapping the same fight, including how data center growth and household power rates interact on regional grids.

On the 13-state PJM grid, which includes northern Virginia’s data-center alley, the independent market monitor Monitoring Analytics put average wholesale power at $136.53 per megawatt-hour in the first quarter of 2026, against $77.78 in the same quarter of 2025. That is a wholesale figure, not a household line item, and wholesale is only part of a residential bill. It is still the kind of print a cartoonist can steal.

The U.S. Energy Information Administration estimated that residential customers would pay an average of 18.2 cents per kilowatt-hour in 2026, a nearly 5 percent increase from 2025, with the largest rises along the East Coast. A Congressional Research Service review notes that past data-center growth did not always lift retail rates, and that Lawrence Berkeley National Laboratory found some high-growth states saw prices fall from 2019 through 2025. The same review says data-center demand appears to be one of several factors in PJM price increases in 2025 and 2026.

WHAT THE RATE NUMBERS ACTUALLY SAY

Those modeled ranges are forecasts, not bills already mailed. They still explain why county meetings about substations, water, and tax abatements have become a national argument. New York passed a one-year pause on large new data centers in 2026. Seattle put in a temporary freeze. Virginia created a new rate class for large customers, effective at the start of 2027. Separate engineering warnings, including that some gas-fired AI campuses may not last a decade, sit on that same grid, not on a foreign message board.

Why the Campaigns Never Reached Real Audiences

OpenAI scored Tech and Tariffs as Category One on its Breakout Scale, meaning the activity stayed on one platform with no evidence it jumped to a real audience. Most of the social posts it found drew little or no observable engagement. The company said it found no evidence of meaningful breakout beyond the operation’s own activity, and no sign that the false ChatGPT-theft claims moved past X or into authentic high-reach accounts.

WHAT OPENAI SAYS IT SAW

• Reach: Category One, one platform, no evidence of breakout.
• Engagement: Little or no observable engagement on most identified posts.
• Theft claims: Entirely false, and not amplified by authentic high-reach accounts.
• Duration: Late 2025 into early 2026, then the accounts were banned.

That failure is not a mystery. The posts leaned on new accounts with few followers. X suspended a share of them. The comics restated stories that U.S. papers had already printed. People who already hate a local project did not need a VPN-written caption to show up at a zoning hearing, and people who want the tax base did not change their minds because a cartoon told them to.

The operators still revealed their working theory. They uploaded a report on how to keep accounts alive, how to mix lifestyle filler with current-affairs posts, and how to use Facebook groups, hashtags, ads, and reporting tools without looking coordinated. They asked ChatGPT to sketch an AI system that would scrape “harmful” posts by “key persons,” store logs, download videos, and send risk alerts. The model returned a general 500-word note on data storage and would not help them collect the data.

China Budgeted for the Campuses It Tried to Stall

The same week OpenAI published the bans, Moolenaar, chairman of the House Select Committee on the Chinese Communist Party, said China was pursuing a five-year $295 billion plan to build data centers across China. His statement put that figure next to the ChatGPT takedown on purpose. In his telling, Beijing wants Americans to fight over U.S. campuses while it pours money into its own.

Cotton, the Senate intelligence chairman, struck the same note in a June 10 letter to Acting Attorney General Todd Blanche. He asked the Justice Department to investigate foreign influence aimed at the U.S. AI buildout, and he wrote that the Chinese government promotes its own AI ambitions at home, where public sentiment is overwhelmingly pro-technology. He said Americans have valid concerns about energy costs and natural resources, and that those concerns can be used by an adversary that wants to slow U.S. compute.

OpenAI drew a longer parallel. In 2022, the Spamouflage, or DRAGONBRIDGE, network used inauthentic accounts to go after Lynas Rare Earths over a planned Texas plant, then turned on other North American rare-earth projects after they announced new capacity. The Australian Strategic Policy Institute and Mandiant both assessed that work as an attempt to bruise competitors to China’s hold on the industry. Tech and Tariffs, OpenAI said, looked like a similar tactic aimed at a U.S. company in a sector where the United States still leads, and the timing sat next to the extra 100 percent tariff and the Party’s AI+ push.

Moolenaar also said Chinese state-run media has published English-language articles that try to split Americans over data-center construction, and that other elements in China are trying the same thing by more disguised methods. He said there are legitimate questions about data centers, that Americans deserve answers, and that his committee will keep looking at malign influence in that debate. He asked people in the fight to treat one another with respect.

A Failed Op Still Colors a Local Fight

The ads did not convert anyone OpenAI could measure. They still handed Washington a clean exhibit. Once a China-linked shop has been caught writing power-bill comics, every local complaint about a substation can be waved off as a talking point that already appeared in a threat report. That is useful to companies trying to pour concrete. It is rough on the households looking at the 18.2-cent rate.

Treating every rate protest as a foreign product is a gift to the people who want the campuses built on the current terms. It is also a gift to the operators, who do not need a viral post if they can make neighbors distrust each other’s motives. OpenAI said foreign influence shops have long latched onto local issues and sincerely held beliefs. In this case they hid who they were and tried to walk into an American argument about the country’s AI capacity.

By September 2026 the June takedown was still being pulled into fights over who counts as a serious critic of data-center siting. The clusters are gone. The capacity auctions are not. The next cartoon will not need ChatGPT if the bills keep moving on their own.

Frequently Asked Questions

Did OpenAI Say the Chinese Government Ran the ChatGPT Campaigns?

No. OpenAI said both clusters likely originated in China and used VPNs because it blocks access from there. It tied Data Center Bandwagon to a private technology firm working for provincial-level government clients and said that pattern fits a commercial trade around public-opinion work. For Tech and Tariffs it flagged public-security phrasing and a user who called the accounts a water army, then said it could not establish a precise institutional affiliation.

What Does Category One on OpenAI’s Breakout Scale Mean?

Category One means the activity stayed on a single platform and OpenAI found no evidence it broke out to a genuine audience. That is the grade it gave Tech and Tariffs. Across both clusters it reported little or no observable engagement on most posts and no evidence of meaningful breakout beyond the operation’s own activity, including on the false ChatGPT data-theft claims.

Were the Fake ChatGPT Data-Theft Posts Written by the Model?

OpenAI said they were not. Beginning in late 2025, a set of likely inauthentic X accounts posted variations of a claim that ChatGPT user data had been compromised and that the theft had damaged their lives. Those accounts had few or no followers and only a few posts. OpenAI found they interacted with the Tech and Tariffs accounts, including quote-posts of the same unrelated tweet within hours, and said X suspended them independently of OpenAI’s own review.

What Older Playbook Did OpenAI Compare the Tariff Cartoons To?

It compared them to 2022 campaigns against North American rare-earth projects, when inauthentic accounts targeted Lynas Rare Earths over a Texas processing plant and later went after Appia Rare Earths & Uranium in Canada and USA Rare Earth after new production plans. OpenAI cited assessments by the Australian Strategic Policy Institute and Mandiant that those drives were meant to harm competitors to China’s position in rare earths, and it said Tech and Tariffs used similar tactics against a leading U.S. AI company, without success.

Did ChatGPT Help the Operators Attack Chinese Dissidents?

Not in the cases OpenAI described. The Data Center Bandwagon cluster asked for short insults aimed at Li Ying (Teacher Li) and at commentators including Lu Yiheng, Xu Chi, and the X account @SydneyDaddy1, and it tried to script messages that would push a YouTuber to have a former Chinese police officer talk about U.S. policy failures. OpenAI said its models refused to generate inflammatory or personal attacks against Li, a person it had already flagged as a target of harassment tied to Chinese law enforcement in an earlier threat report.