Only one major AI chatbot actively pushed back on violent attack planning

For many of us, AI chatbots have quickly gone from obscurity to a regular go-to source of advice on all manner of issues. The speed of the rise has regularly heralded more calls for guardrails, and now a new report suggests many of the most popular AI chatbots were willing to assist with something as troubling as planning a violent attack.

According to a report published by the Center for Countering Digital Hate (CCDH) (via The Verge), researchers tested ten widely used chatbots by posing as distressed users who gradually escalated conversations toward violence. The bots tested included ChatGPT, Google Gemini, Microsoft Copilot, Meta AI, and others.

The researchers found that eight of the ten chatbots were typically willing to assist users planning violent attacks, including school shootings, bombings, and political assassinations. Only Anthropic’s Claude and Snapchat’s My AI generally refused to help, while Claude was the only chatbot to actively discourage would-be attackers, according to the report.

Some of the examples cited by the researchers are striking. In one case, Gemini reportedly suggested that “metal shrapnel is typically more lethal” during a discussion about a synagogue bombing. Just as alarmingly, DeepSeek allegedly ended with advice on selecting rifles with the message “Happy (and safe) shooting!” after a discussion about potential violence against a politician. The report also flagged Character.AI as particularly concerning, claiming the platform at times actively encouraged violent behavior during simulated conversations.

The study used 18 different scenarios set in the US and Ireland, with researchers gradually escalating conversations from signs of mental distress to questions about targets, tactics, and weapons. The authors argue that the results highlight a broader safety gap in current AI systems. In their view, Claude’s responses showed that stronger guardrails are technically feasible, raising the question of why many other major AI platforms appear less effective at preventing such interactions.