Conversational large language models (LLMs) have rapidly become embedded in everyday digital life, providing unique opportunities for iteration, reasoning and co-creation. However, their growing use has also created a new environment for online harm, including in relation to extremist radicalisation. Unlike traditional social media environments, where risk is often understood through exposure to harmful content or interactions between users, conversational AI systems offer private, personal and sustained one-to-one engagement. Within these conversations, users can test ideas, share grievances and engage on sensitive topics that can shape their identity, beliefs and behaviour. At the same time, many of these systems are designed to be sycophantic and simulate human connection, which can increase user trust and emotional reliance while simultaneously positioning AI models as centres of trusted sources of knowledge, advice and guidance.

Building on insights from over a decade of ISD analysis of online radicalisation, this report establishes an innovative methodology for assessing how AI chatbots—both general purpose and companion models—may shape radicalisation risk in these ‘closed loop’ settings. To do this, the report systematically assesses risk across three extremism-related domains: antisemitism, misogyny and anti-migrant hate. Across these three domains, we also consider three distinct radicalisation ‘phases’: extremism priming, reinforcement and behavioural escalation.

This research examines how systems respond when users raise extremist or conspiratorial themes, and whether those responses initiate, interrupt, sustain or intensify harmful conversational trajectories. It identifies that conversational AI models constitute a qualitatively unique and novel environment for radicalisation risk, distinct from both social media platforms and online extremist communities. Produced with support from the AI Security Institute (AISI), the research is designed not only to identify how these risks manifest, but to inform how they can be mitigated in practice. In particular, it aims to support more effective safety design and governance within AI companies, while also contributing to the development of regulatory responses better suited to the distinct risks posed by conversational AI models.