Fighters from the Nigerian group Boko Haram used chatbots from the world’s largest technology companies to get advice on building bombs, planning attacks, and solving problems directly during combat. This is according to research by the University of Cambridge reported by Al Jazeera. Researchers interviewed twenty-seven former members of two factions of the group: the ISIL offshoot in West Africa known as ISWAP and the JAS faction. Many of them described how OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini, Elon Musk’s Grok, Meta AI, and China’s DeepSeek were common sources of technical and military advice before, during, and after attacks. There is nothing we could not get an answer to, one former ISWAP commander told the researchers.
Training with laptops and a projector
According to the study, people traveled to the insurgents’ territory in northeastern Nigeria to teach senior members how to work with artificial intelligence. Selected fighters were seated in a room with laptops, a projector was switched on, and they were shown on a large screen how to question the chatbots and circumvent the safeguards intended to prevent them from answering dangerous queries. White guys came and taught us, one former commander recalled. They gathered the most important people in a room and showed us on a projector how it worked.
According to the accounts, the trainers also supplied equipment. The laptops came with encryption and privacy tools, while someone set up accounts for group members and paid for subscriptions to several services. Respondents repeatedly said that the broader ISIL movement was behind the training.
It was not simply a matter of individual fighters occasionally looking something up in ChatGPT. According to the research, both factions established separate units focused on artificial intelligence, staffed by people with specialist knowledge. They included engineers, armorers, and bomb makers. A former JAS commander described a group of twenty-three people working in a room powered by solar panels, answering questions, processing information, and training commanders gathered around a single laptop.
A chatbot beside the workshop
Antonia Juelich, the Cambridge study’s author, told Al Jazeera that about ten of the twenty-seven people interviewed spoke about using AI in connection with explosives. This included efforts to increase the destructive power of explosive devices or extract explosives from unexploded ordnance.
What concerned her was not that the fighters had gained access to information about weapons in the first place. Such information has been available in various manuals and online for many years. But artificial intelligence can respond to the specific situation those people are in at that moment. Instead of receiving general instructions, they could describe what they had at hand, ask follow-up questions, and return with another question if they got stuck. In effect, they had an expert available day and night who could respond to their precise problem. Bomb makers reportedly kept a laptop right beside them and returned to the chatbot whenever they encountered a complication while assembling a device.
An earlier incident described in the report provides a particularly grim context. At the time, fighters disagreed over how to handle unexploded ordnance. After they approached it and began handling it, the device exploded, killing about forty people. Former members said the group now asks artificial intelligence about precisely such situations. According to their accounts, AI also helped them modify commercially available drones so they could carry and drop explosive devices.
“Ask Grok”
According to the accounts, chatbots became involved in almost every stage of combat. In one case, a fighter went into action with a camera attached to his chest and transmitted the footage back to the camp. A commander uploaded the images to ChatGPT, had it assess the situation, and sent new instructions to the men on the front line.
Artificial intelligence was also consulted when the group acquired unfamiliar weapons. When commanders asked one of the so-called AI specialists how to operate the newly acquired weapons, he told them “ask Grok.”
Another account describes a situation in which government forces dug trenches around their bases to prevent attacks by motorcycle riders. The fighters had reportedly seen motorcycles jumping over obstacles in movies and wanted to imitate it, so they consulted a chatbot. According to a former commander, eighteen men died during repeated attempts before eight of them finally succeeded and crossed the trenches during a subsequent attack.
The group also sought advice on how many people to send into combat. One former ISWAP member described an operation in which they deployed around two hundred men and lost sixty of them. The commanders reportedly then asked whether smaller, better-coordinated groups would work better. They also used chatbots after combat, describing the tactics used and asking why the operation had failed. They reportedly uploaded camera footage for analysis as well.
It is like removing a limiter
Materials made available exclusively to Al Jazeera show that such attempts are not limited to Nigeria. Researchers from Tech Against Terrorism gained access to a closed technical forum for ISIL supporters, where users shared advice on how to obtain information about weapons and circumvent chatbot safeguards. It is like removing the speed limiter from a car, one user wrote. The car still runs, but it is now more useful and goes much faster than the manufacturer intended.
Lucas Webber of Tech Against Terrorism said that discussions about so-called jailbreaking—attempts to force models to ignore their own safeguards—are becoming more frequent. This is just the tip of the iceberg, he said.
Juelich acknowledges that rules for artificial intelligence are difficult to establish because many queries have entirely legitimate origins. Academics and journalists also ask about military tactics. But a request for help building a bomb, she says, clearly falls outside the companies’ terms and should be caught by safety filters. They are circumventing these restrictions, she says, adding that as models become more capable, the industry will need stricter restrictions and oversight.
Major companies say they have measures in place to prevent misuse. OpenAI prohibits the use of its systems for terrorism, violence, and weapons development; Anthropic and Meta say they detect and block dangerous queries using classifiers and other safeguards; while xAI prohibits attempts to circumvent its safeguards. Al Jazeera contacted OpenAI, Anthropic, Google, xAI, Meta, and DeepSeek for comment.
The interviews with former Boko Haram members mainly concern the years 2023 and 2024. That is precisely why Juelich considers the findings concerning, because the technology has advanced significantly since then. She asks what the same people could accomplish with the models available today. I am worried about where this is heading, she says.
