Cambridge Study Reports Boko Haram Factions Used AI Chatbots for Operational Support

A University of Cambridge report presents interview-based evidence that the two Boko Haram factions—Islamic State West Africa Province (ISWAP) and Jama’at Ahl as-Sunnah lid-Da’wah wa’l-Jihad (JAS)—reportedly incorporated commercially available AI chatbots into combat and day-to-day operations. The central implication is not that AI alone created a new capability, but that broadly accessible systems may help organized violent groups obtain technical assistance more quickly and with less dependence on scarce specialists. 423

What the study examined

The research drew on 57 in-person interviews conducted in northeast Nigeria in 2025 and 2026 with 27 former members, including mid-ranking commanders and technical specialists. The report characterizes its findings as on-the-ground evidence of AI use by an active terrorist organization. 4

Interviewees identified six systems—ChatGPT, Claude, Gemini, Grok, Meta AI, and DeepSeek—as tools used by the factions. Their accounts described use across combat and routine organizational activity rather than a single platform or isolated experiment. 23

Reported roles for AI chatbots

According to the report, former members said the groups used AI systems to support attack planning, explosive-device design, weapons servicing and troubleshooting, and operational security. Accounts also described technical and tactical advice in the course of operations. 423

The meaningful change, if the accounts are accurate, is organizational: a chatbot can serve as a fast, repeatable source of structured information. That could reduce reliance on specialist expertise and shorten some problem-solving cycles. The available evidence does not, however, establish that AI enabled any particular attack or measure how much it improved real-world outcomes.

Reported specialist units and controlled access

The report says both factions developed dedicated AI capabilities under senior leadership. Former members described trusted technical personnel—such as bomb-makers, weapons specialists, and engineers—using the systems, interpreting answers, and distributing relevant guidance through the organization. 4

That reported structure matters because it differs from a picture of individual fighters casually experimenting with a chatbot. Access was reportedly controlled for security reasons, with specialized users acting as intermediaries between commercial systems and the wider group. 4

The study also reports that knowledge and training circulated through transnational jihadist networks. Its executive summary says Islamic State-linked networks helped spread relevant know-how, alongside access to devices, connectivity, and accounts. 4

Safeguards were reportedly treated as obstacles, not final barriers

Former members described attempts to evade model safeguards by concealing harmful intent or using prompt-based circumvention techniques. They also described using more than one provider, allowing a refusal or account action on one service to be offset by trying another. 423

These claims should not be read as a test of today’s safeguards. The interviews describe past experiences, and the study does not demonstrate that the same requests would succeed against later model versions or current enforcement systems. Still, the accounts illustrate a core safety problem: protections applied by one provider can be less effective when comparable capabilities are widely available across multiple services.

Why the findings matter for AI safety and security

The study moves the debate beyond the familiar concern that extremists use digital platforms for propaganda, recruitment, or distribution. Its interviewees instead described AI as a component of operational support: a tool for technical questions, planning, and organizational problem-solving. 423

That raises difficult questions for AI companies and policymakers. Platform policies can prohibit weapons development and violent wrongdoing, but enforcement must contend with ambiguous, dual-use requests; users who hide their intent; and an ecosystem in which people can switch tools. The report argues that the accessibility of frontier systems, combined with networks that can share methods and expertise, increases the urgency of that challenge. 4

Important limits on the evidence

The report is serious field research, but its evidentiary boundaries are important. It is based principally on retrospective accounts from 27 former members, not direct observation of every claimed use or independent forensic attribution of attacks to a particular model. 4

Its strongest conclusion is therefore narrower than some headlines suggest: former members described systematic and organized adoption of commercial AI tools by ISWAP and JAS. The interviews provide a warning about plausible operational misuse and the limits of platform-level safeguards, but they do not provide a quantified causal measure of AI’s effect on violence.

For developers, providers, and policymakers, the lesson is clear: preventing misuse cannot rest solely on a written prohibition or a single refusal. It requires ongoing evaluation of real-world abuse patterns, stronger detection and response, and coordination across services where misuse can migrate. 4