Intelligent Systems, Vulnerable Minds: A Framework for Radicalization to Violence in the Age of AI
AI does not create radicalization out of nothing — but recommender systems, generative AI and bot swarms amplify known psychological mechanisms at each stage of the pathway. We map that interaction in a four-stage framework, from first exposure to (rarely) violent action.
The paper, in plain language
Two literatures that had not met
Radicalization research has strong process models; AI research understands the systems. This review — written with colleagues including Milan Obaidi and Anton Gollwitzer — brings them together, connecting stage models of radicalization with what recommender algorithms, generative AI and coordinated bot networks actually do, and with the psychological vulnerability factors the field has established: social isolation, identity insecurity, perceived injustice, and the need for clarity and strong group belonging.
Four stages
The framework traces the pathway through four stages. Exposure: engagement-maximising recommendation and virality features create the first contact with extreme content. Reinforcement: filter bubbles and group recommendations exploit cognitive biases until an echo chamber hardens belief. Group integration: ideologically homogeneous clusters, bot swarms simulating consensus, and AI companions that validate extreme views supply the belonging that radicalization runs on. And rarely — the crucial word — violent extremist action. The pathway is not linear, and it is deeply personal.
Why it matters
The point is not that algorithms are ideologues; they exploit vulnerabilities without any ideological intent, simply by maximising engagement. That is precisely why governance can work: each stage names concrete levers, from recommendation design to synthetic-consensus detection. The framework connects my extremism strand to the AI-swarms work in Science, and it is core to the agenda of the Center for Democracy & Information Integrity, which I co-direct.