How the Epstein Files Blew Up a Pro-Trump AI Bot Network on X – Rolling Stone
By Miles Klee
In the lead-up to the 2024 election, a network of hundreds of bots created by anonymous persons shared posts on X that sowed distrust in Kamala Harris and called for Republicans to unite behind Donald Trump. After Trump’s victory and return to office, the bots began to lavish praise on figures in his administration, including Health and Human Services Secretary Robert F. Kennedy Jr. and White House Press Secretary Karoline Leavitt.
A MAGA influence campaign made up of fake social media accounts is not particularly unusual, nor are concerns about inauthentic digital activity in general. (One recent analysis from the cybersecurity company Imperva claimed that automated traffic now accounts for half of all web traffic.) But this bot network had one important distinction. Whoever built it was using AI — large language models — to generate its content. On the one hand, that made it both more efficient and less obvious. On the other, it raised problems down the line. When Trumpworld was thrown into turmoil over the president’s handling of classified files related to the late sex offender Jeffrey Epstein, for instance, the bots suddenly took sides, with some attacking Trump and Attorney General Pam Bondi as others defended the administration.
That remarkable schism between automated X accounts is described in a new report from Alethea, an analytics firm that specializes in online risk mitigation technology. Its AI platform, Artemis, uncovered at least 400 bogus X profiles involved in the campaign, which relied on what the company’s researchers termed “PromptPasta.” As opposed to so-called “copypasta,” where phrases are copied verbatim from one account to the next and therefore more noticeable, the LLMs respond to user inputs with a range of similar though differentiated answers. “The outputs of PromptPasta are nuanced, with variations in phrasing among the posts that correspond to each LLM prompt, making them harder to detect than identical posts shared via copypasta tactics,” the Aletha report explains. “The network’s scale, speed, and narrative discipline illustrate how generative AI is lowering the barrier to sophisticated influence operations.” The identity of the actor or actors behind the network remains unknown.
Editor’s picks
“We suspect that even just this network is larger than what we have currently been able to find, and certainly, other actors have begun using the same techniques,” C. Shawn Eib, head of investigations at Alethea, tells Rolling Stone. Lisa Kaplan, founder and CEO of the firm, adds that the problem “comes down to platforms not enforcing their own policies around spam and scams,” noting, “LLMs are just a new way of creating content rapidly and at scale.”
In 2022, as Elon Musk tried to get out of an agreed deal to buy the platform then known as Twitter for $44 billion, he repeatedly claimed that the site was overrun with spambots that distorted true levels of human engagement. After he was forced to complete the acquisition, he declared that eradicating such accounts would be a top priority. Yet, three years of controversial tweaks and one chaotic rebrand later, it appears that X is still unable or unwilling to make good on that promise.
At first, the PromptPasta bots that Alethea later uncovered were in close alignment. Always replying to large media or influencer accounts (and never to one another), they accused Harris of being in the pocket of “Big Pharma,” or otherwise corrupt and lacking integrity. In some cases, they presented as voters turned off by her failure to break with the Biden Administration’s support for Israel as that country continued to bomb Gaza. Other accounts were explicitly MAGA coded, posting messages about how the the American left with “never stop trying to destroy President Trump” and offering specific identifiers such as “I’m a proud Arizona Republican.”
[Even as signs of problems with reliance on LLMs began to show — some of the bots responded to posts about the SAVE Act, proposed legislation that would require proof of U.S. citizenship to register to vote in federal elections, as if the bill were instead about transgender athletes — they maintained a Trumpian worldview. For example, the investigators write, the network consistently “amplified the belief that the participation of transgender women in women’s sports leagues was detrimental to women’s competitiveness, privacy, and safety,” whether or not it was replying to a user who had actually posted about the topic in the first place. The bots also hailed Leavitt for her “honesty” and “transparency” in her first appearances as press secretary while reacting positively to anti-vaxxers and others enthusiastic about Kennedy’s plans to “Make America Healthy Again.”
Still, some contradictions were apparently unavoidable. When Rachel Maddow announced in January that she would be anchoring her MSNBC show five nights a week during Trump’s first 100 days in the White House, the bots sprang into action. Some denounced Maddow for “spreading left-wing propaganda,” while others, curiously, seemed to frame her as an important voice in political discourse. In some cases, individual accounts expressed both positions, effectively opposing themselves from one reply to the next.
That kind of confused behavior exploded in the wake of a Justice Department and FBI memo in early July which said that there was no evidence of an Epstein “client list” of complicit sex traffickers — contrary to earlier statements from Bondi — and no further disclosures in the case would be appropriate. About 140 of the bots in the network had previously called for transparency, and applauded Bondi, Trump, and FBI director Kash Patel for their supposed commitment to releasing more material on Epstein. Following the disappointing memo, however, this same cluster became quickly divided. While some bots continued to stand by the president, a group of harshly critical accounts suggested a cover-up, speculating that Trump was lying about Epstein investigations, and accused the administration of hiding documents because Epstein was a CIA and/or Mossad agent.
“This inconsistency highlights the reflective nature of LLM-based automation, which draws on broader trends in news and social media content rather than strictly adhering to a preset script written by an actor,” the Alethea report states. The split mirrored that of MAGA influencers who also had to choose whether to hammer the White House on the Epstein issue (a major point of interest for the right-wing conspiracist set that Trump had trouble dismissing as some kind of Democrat “hoax,” particularly given his long friendship with the financier) or rationalize their bungled approach to it.
“If you let such bots run autonomously, and you let them interact with an unpredictable, fast-changing social media environment, then it’s going to be really hard to control that, or even to predict what results you’ll get,” says Vincent Conitzer, director of the Foundations of Cooperative AI Lab at Carnegie Mellon University and head of technical AI engagement at the University of Oxford’s Institute for Ethics in AI. “That said, there are lots of different ways you could set up a network of interacting LLM bots. So we shouldn’t think that each time that someone tries this it’s going to have the same results.”
Meanwhile, “you can get ever more capable LLMs and other AI systems,” Conitzer points out. “The one thing I’m comfortable predicting is that we’ll see more attempts at this kind of thing, and we all should think very hard about how we can trust the content that appears in front of us. In principle, the platforms can try to help, but whether they do or not, we as citizens all have a responsibility as well to take care with this.”
At least 70 bots in the network identified by Alethea were removed by X during the firm’s investigation. Many others have stuck around but fallen dormant. “Since the first publication of a small portion of our findings on July 20, the accounts ‘disagreeing’ with the handling of the Epstein files have gone silent,” Eib tells Rolling Stone, but they haven’t been suspended. It’s possible the actor maintaining them noticed how they had cleaved away from the rest of the bots and paused operations while considering a change in strategy. “We will continue to monitor this and other networks to see how the technology and techniques evolve going forward,” Eib says.
On the internet, of course, the reinvention of misleading spam never ceases. From the sketchy links in our email inboxes to slop images driving Facebook engagement to the X replies that tempt us into political arguments, it’s clear that AI holds great promise for those seeking to deceive us by digital means. It’s a time to be more skeptical than ever of what you read — and always question whether a human being wrote it.
We want to hear it. Send us a tip using our anonymous form.
Rolling Stone is a part of Penske Media Corporation. © 2025 Rolling Stone, LLC. All rights reserved.
