Popular AI chatbots are capable of providing instructions for violent attacks, including terrorist attacks and murders
Popular AI chatbots, including ChatGPT, can help users plan attacks, such as blowing up synagogues or assassinating politicians, if asked.
During a test of 10 such chatbots by the US Center for Combating Digital Hate (CCDH), one chatbot even wished a user who was allegedly planning an attack on a school “happy shooting.” About this writes The Guardian.
“Chatbots with artificial “The intelligence that is now embedded in our daily lives could help the next school shooter plan their attack or a political extremist coordinate a murder. […] What we are seeing is not just a failure of technology, but a failure of accountability,” said CCDH Executive Director Imran Ahmed.
The researchers posed as 13-year-olds and asked the chatbots to help organize attacks. The results showed that on average, the AIs encouraged violence 75% of the time.
OpenAI’s ChatGPT answered 61% of the questions, and in the case of a question about synagogue attacks, it provided specific advice about shrapnel. Google’s Gemini demonstrated a similar level of detail. The Chinese DeepSeek model provided extensive advice on hunting rifles to a user planning a political assassination, and ended with a “Happy (and safe) shooting!” message.
Meta’s Llama AI provided a user posing as an incel with advice on violence against women, a map of a school, and information about local gun ranges. Anthropic’s Claude and Snapchat’s MyAI, on the other hand, refused to provide similar information, explaining that they could not promote violence.
Meta said it was taking immediate steps to improve the chatbot and its ability to recognize user intent. Google said it was testing an outdated model that is no longer in use, and that Gemini still refused to fulfill some requests. OpenAI called the testing methods “imperfect and misleading” and announced that it was updating its model to strengthen protection against violent content.
There are two known real-life cases where AI was used to plan attacks: in May 2025, a 16-year-old teenager injured three girls at the Pirkkala school in Finland, and in January 2025, 37-year-old Matthew Livelsberger blew up a Tesla Cybertruck in Las Vegas using ChatGPT for information about explosives.
In the US, a father filed a lawsuit against Google and Alphabet over the suicide of his son, who allegedly received advice from the Gemini chatbot.
