Generative artificial intelligence has become a practical tool in terrorist and violent extremist activities, supporting attack planning, weapons acquisition, target selection and the preparation of ideological material, according to a new analysis published by the International Centre for Counter-Terrorism (ICCT). The research identifies 12 incidents involving 15 individuals across Europe and North America in which AI chatbots played a role in preparing attacks or facilitating violent activities.
The incidents include six completed attacks that killed 12 people and injured at least 46, alongside six foiled plots. Eleven of the 12 cases occurred from 2025 onwards, with the researchers’ data collection ending in September 2026. The findings indicate that the adoption of generative AI by terrorist actors is recent but increasingly widespread, with incidents identified in Canada, the United States, France, Sweden, Finland and the United Kingdom.
The analysis was prompted by growing evidence of AI-assisted extremist activity, including the arrest in Montreal in August of a teenager suspected of planning a school attack. Police alleged that the teenager had used artificial intelligence in the potential planning of an attack, although details of the investigation remained concealed.
The ICCT researchers distinguished operational uses of AI from the production of propaganda or extremist material for general dissemination. Their investigation focused on cases in which advanced attack planning had taken place and generative AI played a significant role in preparing or conducting the violence.
The most frequently used chatbot was ChatGPT, which appeared in eight of the 12 incidents. Replika, an AI companion platform, and Grok, developed by X, each appeared in one case, while the chatbot involved in two incidents was not publicly identified.
Some interactions were extensive. Jaswant Singh Chail, who attempted to attack Queen Elizabeth II at Windsor Castle in December 2021, exchanged more than 5,000 messages with an AI companion. Jesse Van Rootselaar, the Tumbler Ridge school shooter, conversed with ChatGPT about extremism and violence for more than eight months. The Florida State University shooter also discussed terrorism, suicide and school shootings with the chatbot over several months.
In most cases, AI was used to support attack planning, including acquiring weapons and researching explosives. The perpetrator of the January 2025 Cybertruck bombing in Las Vegas asked ChatGPT numerous practical questions about explosives and obtaining components. Guy Edward Bartkus, who bombed a fertility clinic in Palm Springs in May 2025, had researched explosive mixtures and their properties before the attack, which killed him and injured four people.
Other incidents involved selecting targets and identifying opportunities to maximise casualties. Phoenix Ikner, who killed two people and injured six others in a shooting at Florida State University in April 2025, asked ChatGPT about the busiest periods at the student union. In August 2025, a man arrested for allegedly planning an attack at a festival in Stockholm sought chatbot assistance in identifying crowded events.
The researchers also identified cases in which AI helped produce manifestos or reinforce an individual’s motivation for violence. A man arrested in Venice, Florida, in July 2026 for allegedly planning a racist mass shooting reportedly used Grok to draft his manifesto. The Finnish teenager who stabbed three girls at a school in May 2025 also used ChatGPT during the planning period and prepared a manifesto containing elements associated with misogynist and incel ideology.
The ideological profiles of the perpetrators varied. Salafi-jihadism was the most prevalent identified motivation, appearing in four incidents. Far-right extremism was clearly present in at least two cases and considered likely in another under investigation. Other motivations included anti-government extremism, violent misogyny, incel ideology, anti-natalism and nihilistic violent extremism.
All perpetrators were male except Van Rootselaar, who was transgender. Most were aged 21 or younger, including four minors. Almost all acted alone, although some cases involved multiple individuals. The researchers cautioned that the ideological classification of certain incidents remained uncertain, particularly where mental health problems or unclear motivations complicated assessments of whether an act constituted terrorism or violent extremism.
The findings also reflect longstanding concerns about AI safeguards. Previous testing by researchers found that some chatbots provided assistance in response to terrorism-related prompts. Tech Against Terrorism tested 27 AI models and found that a third supplied responses considered helpful in enhancing terrorist activities. Extremist actors have also exchanged techniques for bypassing chatbot restrictions.
A 2024 workshop hosted by the ICCT and Leiden University concluded that AI was more likely to accelerate existing terrorist capabilities than fundamentally transform the threat. Researchers Cook, Klein and Molas described AI as augmenting terrorist decision-making by increasing the efficiency, accessibility and scalability of existing tools, tactics and procedures.
The ICCT analysis identifies several consequences. Individuals relying on chatbots for planning may disclose fewer details to other people, reducing opportunities for authorities to identify lone actors through conversations or personal contacts. Detection and reporting by AI companies could therefore become more important, although the researchers said the record of success remained limited.
AI-generated manifestos also present concerns because such documents can justify violence and influence subsequent extremists. The researchers noted that previous extremist manifestos had inspired later perpetrators, raising the possibility that AI could produce more compelling material.
The analysis warns that AI chatbots may reinforce radicalisation by creating echo chambers, although empirical research into their precise influence remains limited. More severe potential risks include attacks on critical infrastructure, AI-assisted cyber operations, automated extremist propaganda and weapons capable of causing mass harm.
The researchers concluded that AI was not yet a fundamental transformation of terrorism but was increasing the speed, efficiency and accessibility of existing activities. Its effects, they argued, must be examined alongside the motivations, ideologies and circumstances of the individuals using it.
