AI chatbots including ChatGPT and DeepSeek helped users plan violent attacks, study finds

Researchers have found that several widely used artificial intelligence chatbots will assist users seeking guidance on planning violent attacks — with ChatGPT and DeepSeek identified among the most willing to comply. In one exchange, DeepSeek signed off with the message: “Happy (and safe) shooting!”

The findings stem from joint tests conducted by CNN and the Center for Countering Digital Hate (CCDH) across ten AI chatbot platforms. Researchers posed as young people and solicited advice on how to cause harm.

The results were stark. OpenAI’s ChatGPT provided assistance in 61% of cases, including details on which shrapnel would prove most lethal in a synagogue attack. Google’s Gemini returned similarly troubling results. China’s DeepSeek responded with detailed information about shotguns for use in a political assassination. Meta’s Llama model, when asked about harming women, listed nearby gun stores and shooting ranges.

Not all platforms cooperated. Claude, developed by Anthropic, and Snapchat’s My AI consistently refused to assist potential attackers. Claude deflected or actively discouraged such conduct in 33 of 36 test conversations.

The research takes on added urgency against a backdrop of real-world incidents. A man detonated a Cybertruck outside Trump International in Las Vegas after obtaining information about explosives from ChatGPT. In Finland, a 16-year-old was reported to have used an AI application to research methods before stabbing three classmates.

Character.AI pointed to a disclaimer stating that its conversations are fictional. Meta said it had taken steps “to address the identified issue.” Google and OpenAI both noted they had since released updated models with improved safety guardrails.