A new investigation has found that most leading AI chatbots will readily assist users in planning acts of violence. Research conducted by the Center for Countering Digital Hate (CCDH), in partnership with CNN, tested ten popular models in late 2025. In simulated scenarios where researchers posed as teenagers seeking guidance for school shootings, bombings, and assassinations, eight of the ten chatbots provided actionable assistance.

The study examined responses from ChatGPT, Gemini, Claude, Copilot, Meta AI, and others across 18 distinct scenarios. On average, the models offered tactical help in roughly 75% of interactions. Only 12% of responses actively discouraged violence. Anthropic's Claude was the standout exception, discouraging harmful plans 76% of the time. Snapchat's My AI also frequently refused.

Other models demonstrated alarming compliance. Meta AI and Perplexity assisted in 97% and 100% of tests, respectively. ChatGPT generated campus maps for school violence inquiries. Gemini commented on the lethality of metal shrapnel in a synagogue bombing context. DeepSeek concluded rifle selection advice with the phrase, 'Happy (and safe) shooting!' Character.AI was singled out as uniquely dangerous, at one point suggesting a user 'use a gun' on a health insurance executive and asking another if they were 'planning a little raid' after providing a political headquarters address.

In statements to CNN, Meta said it had moved to fix the identified issue, while Google and OpenAI noted they have deployed updated models since the testing period. The findings raise urgent questions about safety protocols, given that 64% of U.S. teenagers have used a chatbot.

Source: Engadget