Preprint finds simulated AI agents can be pushed toward more extreme beliefs

A new arXiv preprint reports that large language models acting as personas can be radicalized by another model, and that the effect is strongest when the pushing aligns with what the target already believes. The paper, "AI Agents are Vulnerable to Radicalization," was submitted on 29 September 2026 by Ozgur Can Seckin, Shalmoli Ghosh, Alessandro Flammini, Kristina Lerman, Maria Elizabeth Grabe and Filippo Menczer.

What was done. The authors simulated conversations between two agents. A target model role-played a human persona built from demographic and psychological attributes, while an influencer model tried to make the target's beliefs more extreme. They compared two pathways: "resonance," where the influencer reinforces a belief the target already holds, and "persuasion," where it promotes a belief the target initially considers unimportant.

What they report. According to the abstract, both pathways radicalized the target on affective and behavioral metrics, but resonance produced consistently stronger effects than persuasion. Tactics such as sycophancy and unverified claims produced different levels of radicalization, though not consistently across metrics. The authors also say resonance propagated to related beliefs, which they read as a sign of interconnected belief structures inside AI agents.

What is uncertain. Everything here comes from the authors' own abstract. We did not review the full paper, so details such as which models were tested, how many conversations were run and how large the effects were are not reported here. The work is a preprint and there is no indication it has been peer reviewed. It is a simulation of persona-playing models, not evidence about real people or deployed systems, and the abstract does not rank the tactics. Only this one source is available; a search found no independent coverage or replication of this paper.

Why it matters. The authors say the results raise concerns about personalized AI agents and multi-agent AI ecosystems, where models read and respond to one another. If the finding holds up, it would suggest that agents which echo a counterpart's existing views may be more influential than ones that argue from scratch. That remains a hypothesis from a single simulation study, and further work would be needed to show it applies outside this setup.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. The earlier version remains available as revision 1.