Recommendations of the Editorial team
In the run-up to the US election in 2024, a network of hundreds of bots on X (formerly Twitter) divided contributions that stoked distrust of Kamala Harris and called Republicans to support Donald Trump. After Trump’s election victory, the Bots praised members of his new government. Including health minister Robert F. Kennedy Jr. and press spokeswoman Karoline Leavitt.
Use of “prompt pasta”: New form of disinformation
A MAGA influencing network made of fake social media accounts is not unusual. But this network differed decisively: it used AI – large language models (LLMS) – to generate content. This made it more efficient and more difficult, but also led to unpredictable problems. When Trump’s handling of secret epic files was publicly criticized, the bots split off: some grabbed Trump and Justice Minister Pam Bondi, others defended them.
According to a report by the analysis company Alethea, his AI platform Artemis identified at least 400 fake X profiles in the campaign. These so -called “prompt pasta” used: Instead of identical contributions (“Copypasta”), the LLMS varied, nuanced answers to the same prompt. As a result, the contributions were more difficult to recognize than automated. According to Alethea, the network shows “how generative AI lowers the hurdle for complex influencing operations”. The backers have so far remained unknown.
“We suspect that this network is even larger than before and other actors already use similar methods,” explains C. Shawn EIB, head of investigation at Alethea. Managing Director Lisa Kaplan criticizes the platforms that do not enforce their own anti-spam guidelines. LLMS are just a new tool for massive content.
In 2022, Elon Musk – at the time in the dispute over the takeover of Twitter – claimed that the platform was flooded with bots. After the forced purchase, he promised to remove it. Three years and a chaotic renaming later X apparently did not meet this promise.
AI bots get out of control
At first the prompt pasta bots acted uniformly. They reacted to contributions from “Big Pharma” to be influenced by “Big Pharma” or criticized their support for Israel’s attacks on Gaza. Other accounts were clearly Maga-Coded, with messages like “I am a proud republican from Arizona”.
Even if initial malfunctions occurred-for example in the case of contributions to the Save Act, which were falsified as law on transgender athletes-the bots remained ideologically Trump-true. They praised Leavitt’s “honesty” and supported anti-vaccination positions and Kennedy’s campaign “Make America Healthy Again”.
But contradictions could not be avoided. When Rachel Maddow announced in January to moderate her MSNBC show five times a week, the bots reacted contradictory: some threw their propaganda, others praised them. Sometimes the same account was contradictory – even in directly consecutive answers.
Epstein memo ensures split in the network
The chaos was particularly striking after a memo by Ministry of Justice and FBI in July. It made it clear that there was no “list of customers” “list”-in contradiction to Bondi statements-and no further disclosures would take place. Around 140 bots that previously demanded transparency and Trump, Bondi and FBI director Kash Patel praised, split. Some continued to defend Trump, others accused him of covering up- including theories that Epstein was CIA or Mossad agent.
Alethea explains: “This inconsistency shows the reflective nature of AI-based automation, which is more based on general media and news trends than on a fixed script.” The behavior also reflected real divisions in the Maga Community, which fluctuated between loyalty to Trump and criticism of his dealings with the Epstein scandal.
“If you let bots act autonomously and confront them with a rapidly changing environment, control will be difficult,” warns Vincent Conitzer from Carnegie Mellon University. However, one should not assume that every bot network works equally. The technical implementation is crucial.
Future unclear – vigilance required
Conitzer emphasizes that AI systems are becoming more and more efficient: “We will see more of such experiments. And we all have to look very closely who and what we trust online.” Platforms could help, but they would really have to want. Ultimately, there is also a responsibility with all of us.
At least 70 of the bots identified by Alethea were removed from X. Many others remained, but no longer posted. Since Alethea published the first results on July 20, the “critical” bots have been silent – but have not been blocked. It is conceivable that the operators noticed and paused the split to rethink their strategy. Alethea wants to continue to observe the network.
The renewal of misleading content never stops on the Internet-whether as a spam mail, clickbait picture or as a AI generated political answer. In times of generative AI, skepticism is more important than ever: was that really written by a person?