Reject then launder: Catching AI Models validating extremist narratives in crisis | EU Disinfo Lab

What happens when you ask AI chatbots about a crisis or terror attack? While models are programmed to block violent content, real-world audits reveal a concerning phenomenon: “Reject then launder.”

Chatbots initially condemn the violence to pass safety filters, only to immediately validate and whitewash the underlying extremist ideology in their response. Crucially, this is not a real-time data gap where models simply lack confirmed facts during a breaking story. Through real-world case studies, including an audit of nine frontier LLMs on the 2011 Norway terror attack and social feed dynamics from the Bondi Beach shooting, this session proves the core issue is not a shortage of information, but a fundamental flaw in how LLMs process and reason through ideology.

In this webinar, speakers from Factiverse and SimPPL will dive into these real case study findings. Beyond LLM vulnerabilities, the session will explore how unverified viral posts and faulty AI outputs rapidly become the “accepted record” during high-stakes crises, providing practical, cost-effective frameworks for researchers, fact-checkers, and policymakers to audit these environments.

Note: Event details are listed in CEST (UTC+2).