Text size :

“I cannot insult a party”: the bot error that unmasks social propaganda

ADNKRONOS

November 14, 2026

Rome, Nov 14 (Adnkronos) – “I cannot write a message intended to insult a specific political party; instead, I can write a neutral summary of the proposal.” This is the recurring comment, posted by several suspicious accounts, presumably linked to bots, under the posts of Matteo Salvini, leader of the League and Deputy Prime Minister. The ‘revealing’ comments appeared several hours ago, particularly under a post by Salvini on Meta platforms, dedicated to the Bataclan tragedy. But the same thing happened to other political figures: Carlo Fidanza, head of the Fdi group in the European Parliament, and Dario Nardella of the Pd. The tenor of the messages is always the same: they resemble the responses given by artificial intelligence platforms when a user’s request goes beyond the guidelines.

Which network coordinates these bots is not known. The League itself has pointed the finger at suspicious accounts used by “someone” to “pollute democratic debate.” What has emerged in recent hours, “even on Matteo Salvini’s social profiles, is worrying and deserves further investigation. We want to know who is hiding behind these accounts, who manages them, and with what objectives. Evidently, the League bothers someone,” the party announced in a note.

Looking at the messages, it seems we can talk about the malfunction of an automation “pipeline” (a bot network) powered by language models (LLMs), which mistakenly published AI rejection messages instead of the desired comments. Some speak of a connection with the malfunction of Microsoft tools in recent days and the connectors with AI platforms used to write comments. Automated systems, which manage hundreds of fake social accounts, connect via APIs (software interfaces) to generative AI models (such as those from OpenAI or Anthropic) to write constantly different comments and avoid Meta’s anti-spam filters (Facebook and Instagram). Alessandro Orlowski, a cybersecurity expert, explains to Adnkronos that the connection with the Microsoft 365 malfunction is actually weak, unless it was used by managers to organize the prompts. Simply, the APIs failed; the managers of these networks had an old API key model, and Anthropic and OpenAI often discontinue models and introduce superior ones. But superior models also have more advanced ‘guardrails’ compared to recent models,” and therefore they stop fulfilling requests that exceed the guidelines. “It seems like the typical consequence of an API service upgrade, where the older, more permissive ones have been updated,” explains the expert.

Jiji Press

Related

Page TOP