Boko Haram and ISWAP set up dedicated units to study AI chatbots for attacks, a Cambridge study based on 57 interviews with former fighters found this month.
Researcher Antonia Jülich of the Cambridge Programme on AI Science and Policy conducted the interviews with 27 former members over the past year, and The New York Times reported the findings on Friday. Both the main Boko Haram faction and its splinter group, the Islamic State West Africa Province (ISWAP), organized their own AI teams rather than leaving the tools to individual fighters. In one instance described in the study, members gathered senior commanders in a room and used a projector to walk them through how the chatbots worked.
Former insurgents told Jülich the groups turned to platforms including ChatGPT, Claude, Gemini, Grok and DeepSeek for tasks that ranged from planning attacks to maintaining weapons and tightening operational security. One frequently cited case involved a stalled assault on a military base, where a defensive trench blocked fighters on motorcycles. After watching a stunt in a film, the group asked a chatbot how to modify the bikes to clear a similar gap. “We used AI to learn how to do this,” a former commander told the researcher, and mechanics reportedly adjusted the motorcycles for speed before fighters rehearsed the maneuver.
The study also found that ISIS liaisons had coached Boko Haram commanders on getting around chatbot safety filters, and former members said they routinely framed harmful requests as school assignments or engineering projects to slip past the guardrails. OpenAI said such use breaches its policies, while Google and Anthropic said their systems are built to refuse requests tied to weapons or other dangerous activity.
Jülich cautioned that the group’s use of AI has stayed largely conventional so far, but warned it points to a real risk of extremist groups seeking AI help with chemical or biological weapons down the line. Other researchers who reviewed the findings noted a separate concern: general chatbots mostly surface information that already exists elsewhere, so the bigger long term threat may lie in more specialized life sciences AI tools rather than in ChatGPT or Claude themselves.
