One former ISWAP commander summed it up bluntly: “There is nothing we haven’t received an answer on.” That quote, pulled from a University of Cambridge study, is probably the most damning single line any AI safety team will read this year. According to Daily Post, researchers interviewed 27 former members of Boko Haram’s two main factions, ISWAP and JAS, and found that multiple fighters had used ChatGPT, Claude, Gemini, Grok, Meta AI, and DeepSeek to get operational military and weapons information before, during, and after attacks.
The study’s scope is specific. Researcher Antonia Juelich focused on AI use during 2023 and 2024, and roughly 10 of the 27 interviewees described using these tools in direct connection with explosives. That’s not a fringe finding. Fighters described consulting AI while actually constructing explosives, asking chatbots to troubleshoot based on the specific materials they had available. One former commander compared the technology to “a human robot.” The responses, apparently, were useful enough to keep asking.
What makes this more than a simple misuse story is the organizational layer behind it. The study found that both ISWAP and JAS built dedicated AI units, teams of engineers, weapons specialists, and bomb-makers who were trained to extract information from these systems and route it to senior commanders. One former JAS commander described a 23-member unit operating out of a solar-powered room, analyzing intelligence, answering questions from leadership, and teaching others how to use the tools effectively. This wasn’t ad hoc experimentation. It was structured adoption.
The uses extended beyond bomb-making. Fighters reportedly used AI to recover explosive material from unexploded military ordnance, to modify commercial drones for weapons delivery, and to analyze battlefield footage. In one documented case, images from a frontline engagement were uploaded to ChatGPT, analyzed remotely, and used to issue real-time combat instructions. Failed operations were also reviewed through AI, with fighters describing their tactics and asking why an attack hadn’t worked.
Every major platform named in the study, OpenAI, Google, xAI, Anthropic, Meta, prohibits use of their systems for terrorism and weapons development. But prohibition and enforcement are different things. Juelich noted that fighters were specifically trained to find workarounds for safety filters. Research from Tech Against Terrorism found parallel discussions among ISIL-aligned communities about jailbreaking AI systems for weapons-related queries, and analysts there warned that what’s been documented is likely only a fraction of actual activity.
The broader concern here is timing. The study covers 2023 and 2024, before the current generation of more capable models was widely deployed. Juelich said she worries about where this is heading, and that concern is reasonable. If structured jihadist units were already systematically extracting weapons guidance from these platforms two years ago, the question for AI developers isn’t whether misuse is happening. It’s how much worse it’s gotten since.
