When someone falls down an ideological rabbit hole, standard online moderation has only one answer: block, ban, and filter. But while platforms wipe the keywords, the subculture simply invents a new vocabulary, and the grievance deepens in private.

What if conversational AI didn't just enforce rules, but attempted to interrupt the trajectory toward radicalisation before it turned into harm?

To test whether AI can actually challenge deeply entrenched worldviews, I engineered a multi-agent safety system using synthetic persona engines and deployed it across 180 simulated conversations with radicalised incel profiles. Rather than relying on a single approach, the system systematically pitted distinct intervention strategies (Socratic questioning, direct persuasion, and hybrid tactics) against varying tiers of ideological commitment, from newly joining the subculture to full radicalisation.

In this talk, we’ll step inside the state graph of an AI de-escalation experiment. We’ll break down how synthetic personas simulate subcultural fatalism, how different conversational strategies perform over multi-turn interactions, and what happens when an agent tries to reason with an adversary that rejects its premise.

Most importantly, we’ll confront the empirical and ethical limits of the project: what the simulations revealed about ideological resistance, why conversational compliance isn't the same as de-radicalisation, and what this means for the future of AI safety and human persuasion.

An inside look at using multi-agent state machines, synthetic persona engines, and multi-strategy conversational interventions to probe whether AI can disrupt radicalised trajectories and where algorithmic de-escalation hits an immovable wall.