X said it identified about 200,000 suspected Chinese inauthentic accounts, including 200 that posted about U.S. data centers and energy policy, while OpenAI and outside researchers have reported related activity but no demonstrated network-wide audience impact.3 4 5
By PAI
Sept. 3, 2026
SAN FRANCISCO — X said its Safety team identified a “bot farm” of approximately 200,000 suspected Chinese inauthentic accounts, including 200 that posted about U.S. artificial intelligence and energy policy. The company’s Aug. 28 announcement described the content and said accounts that violate its Authenticity policy are suspended, but it did not provide a data set, an account list, an engagement total or a reach estimate for the network. 3
The announcement has been cited in arguments over opposition to data-center construction. In a Sept. 2 LinkedIn post and a linked article, researcher Renee DiResta said the disclosure should not be used to dismiss domestic concerns about AI infrastructure and argued that independent review of platform data is needed to distinguish an operation’s activity from its effect. 1 2
An earlier LinkedIn post by Keith King linked to Newsweek’s Aug. 28 report and described AI infrastructure as a new front in information warfare. King argued for transparent communication and stronger platform defenses while acknowledging that communities have legitimate reasons to scrutinize electricity, water, land and grid effects. Those are King’s policy views, not measurements of the network’s influence. Newsweek said X had shut down the bot farm, but X’s own announcement did not state that every one of the approximately 200,000 accounts had been suspended. 3 13 14
The available records support separate conclusions: X has described suspected inauthentic activity; OpenAI has documented a separate, likely China-origin cluster that used its tools to create related content; and a Clemson University researcher found no engagement for at least 59 accounts in an accessible sample. Those records do not establish the reach of every account in X’s reported network, whether the two companies identified the same full network, or whether the activity altered U.S. views or policy. 3 4 5
What X disclosed
X’s Global Government Affairs account said the Safety team had investigated “suspected Chinese inauthentic accounts involved in influence operations.” It said approximately 200,000 accounts were in the reported farm and that 200 posted “in a manner that could manipulate a legitimate debate” about American AI and energy policy. According to X, the posts said data centers were raising household electricity prices and straining the grid; some used AI-generated cartoons depicting data-center operators enriching themselves at the public’s expense. 3
The wording is narrower than a finding that China’s government directed the accounts or that the accounts caused opposition to data centers. X characterized the accounts as “suspected Chinese” and said only that the 200 posts could manipulate debate. Its public post did not explain the technical basis for its country assessment, identify the operators, specify the period covered, publish the posts or quantify impressions, followers, reshares, likes or other measures of audience exposure. 3
X said it takes attempts to undermine the “integrity of the global town square” seriously and suspends accounts that violate its Authenticity policy. It did not say in the announcement that every one of the approximately 200,000 accounts had been suspended, nor did it describe the enforcement action taken against a particular account. 3
The distinction between an account count and an impact measure is material. A large count can describe the scale of an alleged network, while engagement and exposure data would bear on how much of its content was seen or amplified. Neither is a direct measurement of a change in public opinion, voting, local permitting or energy policy. X’s announcement supplied the former but not the latter measures. 3
A related OpenAI case
OpenAI reported in June that it had banned a cluster of ChatGPT accounts it assessed as likely originating in China. The company said the operators used simplified-Chinese prompts, asked for English- and Chinese-language outputs, posed online as Americans from different backgrounds and posted content across multiple social-media platforms. OpenAI said the operators likely worked on a social-media operations team at a private Chinese technology company serving provincial-level government clients; that is OpenAI’s assessment, not a public attribution by a U.S. government agency. 4
The OpenAI case included English-language comments and images claiming data centers and AI applications were increasing electricity demand and costs for ordinary Americans. OpenAI said operators used prompts to create comic strips about capacity-auction prices, edit power-market images and produce posts with hashtags including #capacityauction and #datacenters. It also said the actors used its tools for workflow tasks such as formatting data and seeking code to automate logins and interactions across platforms. 4
OpenAI’s report describes operator behavior and its own account-enforcement action. It did not offer a quantified finding that the campaign had persuaded an audience or changed the data-center debate. Its report says the group appeared to be part of a commercial ecosystem aligned with party-state priorities, a formulation that does not identify a government command relationship. 4
DiResta wrote that OpenAI’s and X’s disclosures appeared to concern “some of the same exact network.” Clemson researcher Darren Linvill told Business Insider that the accounts were “very, very clearly connected.” Neither company’s public disclosure, however, supplies account-level evidence sufficient to establish that the reported populations are identical. 2 4 5
The available reach evidence
Business Insider reported that Linvill, co-director of Clemson University’s Media Forensics Hub, used a Clemson platform that retains data from suspended X accounts to search an uncommon hashtag shown in X’s disclosure and to identify accounts that shared images X had flagged. He told the publication that at least 59 of the accounts had zero engagement before suspension. 5
The newspaper reported that the #capacityauction search returned 90 posts, all created in January, and that each identified account posted two or three times before X suspended it. Linvill shared a spreadsheet of the underlying data with the outlet. He said it remained possible that some people saw the posts, but described the sampled activity as unsuccessful. 5
That is evidence about a limited, retrievable set of posts and accounts, rather than a measurement of the reported 200,000-account farm. It supports no conclusion about the audience for the other accounts, the activity that did not use the searched hashtag or whether posts appeared elsewhere. X has not released the information needed to calculate those broader measures in public. 3 5
Energy demand is independently documented
The substantive questions raised in the posts — electricity demand, system capacity and costs — do not depend on the authenticity of the accounts that raised them. The Department of Energy said a Lawrence Berkeley National Laboratory report estimated that U.S. data centers consumed 176 terawatt-hours, or about 4.4% of U.S. electricity, in 2023. The report projected 325 to 580 terawatt-hours and 6.7% to 12% of U.S. electricity by 2028. Those are national estimates and projections, not a finding about a particular household bill or facility. 6
The Energy Information Administration reported in May that commercial electricity sales in Virginia increased by nearly 30 million megawatt-hours between 2019 and 2025. EIA said that growth was largely driven by the state’s concentration of data centers, as well as electric-vehicle adoption and building electrification. It also said PJM Interconnection expects the Dominion zone, which covers Virginia, to have the largest absolute increase in summer peak demand from 2026 through 2030, largely because of data-center load growth. 7
Those findings establish that data centers are a significant component of current and projected power demand in a major U.S. market. They do not establish that data centers alone determine retail prices, grid reliability or local development decisions, because those outcomes also depend on electricity-market rules, generation, transmission, other loads and state regulation. 6 7
Earlier platform transparency
DiResta’s broader argument concerns access to evidence after a platform removes accounts. Twitter, before it became X, said in a 2021 transparency update that it had shared 37 datasets of attributed platform-manipulation campaigns from 17 countries, spanning more than 200 million posts and nine terabytes of media. It said raw data allowed outside researchers to examine narratives and how widely they spread, and announced the Twitter Moderation Research Consortium to provide eligible researchers with more comprehensive data. 8
In a 2022 update, Twitter said it was sharing information-operations data with consortium members, including its assessment of presumptive country of origin based on technical signals, while no longer publicly identifying a specific state or other actor. It said Stanford Internet Observatory, the Australian Strategic Policy Institute and Cazadores de Fake News would publish independent research on 15 data sets. 8
An archive of Stanford Internet Observatory material describes 92 platform-takedown investigations from 2019 through 2024. The archive, maintained by former SIO research manager Shelby Grossman, says it preserves research and analyses after the observatory’s closure and cautions that its AI-assisted compilation may contain errors. The archive nonetheless documents a body of work larger than the “more than 70” analyses DiResta cited. 2 9
The historical practice does not show that X is required to disclose its current investigation or that disclosure would settle attribution or impact. Twitter’s 2021 statement itself said confident attribution is not always possible and that data sharing creates privacy, security and safety considerations. It does show that account removals, platform attribution and independent analysis have been treated as separate steps. 8
State Department reference remains unverified
DiResta’s LinkedIn post said a State Department contractor had a listing for researchers able to analyze foreign propaganda, map bot networks and conduct open-source investigations. The accompanying Substack article corrected an earlier assertion about timing, saying the advertisement had appeared earlier and that a former State Department employee’s post highlighting it came five days after Under Secretary Sarah Rogers’ document release. Neither item names the contractor or links directly to the listing. 1 2
The State Department’s archived Global Engagement Center page says the center closed Dec. 23, 2024. Its former mission included recognizing, understanding, exposing and countering foreign state and nonstate propaganda and disinformation. The department’s current page says the Under Secretary for Public Diplomacy, a post held by Rogers, leads efforts to engage, inform and understand foreign audiences. Neither official page identifies the alleged listing, an award, a successor program or an operational result. 10 11
A separate Justice Department case illustrates why sources of evidence matter. In 2024, prosecutors announced domain seizures and searches of 968 X accounts in an investigation of a Russia-linked, AI-enhanced bot farm; the department published affidavits and said X voluntarily suspended remaining identified accounts. That legal record is different from X’s 2026 public statement about suspected Chinese accounts, which has not been accompanied by comparable public evidence. 12 3
The present record therefore documents a platform allegation and enforcement policy, a related OpenAI account ban, limited third-party evidence about a sample’s engagement and independently measured growth in data-center electricity demand. It does not resolve the identity or reach of the full X network, the relationship between the platform disclosures or the effect of the activity on American opinion and energy policy. 3 4 5 6 7