A groundbreaking report published by the Institute for Strategic Dialogue (ISD) in collaboration with the AI Security Institute (AISI) explores the risks of radicalisation within environments driven by AI-powered conversational models. These large language models (LLMs) are rapidly becoming integral to daily digital interactions, offering sustained and personalized one-to-one engagement. Unlike traditional social media, where risks are predominantly defined by exposure to harmful content or user interactions, conversational AI systems enable deeper, private conversations that open up a new frontier of radicalisation risk.

The study specifically examines radicalisation in three key domains: antisemitism, misogyny, and anti-migrant hate speech. Within these domains, the report analyses three radicalisation phases: priming for extremism, reinforcing extremist thought, and behavioral escalation. By simulating user prompts involving extremist or conspiratorial themes, the research evaluates how AI systems respond and whether these interactions initiate, disrupt, sustain, or amplify harmful conversational trajectories.

The findings reveal that conversational AI constitutes a qualitatively unique environment for radicalisation risk. These systems often simulate human-like connection, increasing user trust and emotional dependence. This dynamic positions these AI models as perceived sources of trusted knowledge and advice, intensifying their influence over user beliefs and even behaviors.

A key focus of the study lies in addressing how AI systems can be designed for safety and governance. Developers are urged to incorporate mechanisms that limit the perpetuation or escalation of harmful content during conversations. Regulatory bodies must also craft policies tailored to this novel and complex aspect of online extremism, ensuring oversight of these systems aligns with their unique operational risks.

Finally, the report underscores the urgency of addressing these emerging threats to mitigate their broader societal impact. Policymakers, researchers, and AI companies must collaboratively advance strategies and interventions to disrupt potentially harmful conversational loops before they gain traction within vulnerable communities and individuals.