Eight of the 10 AI chatbots tested in a joint investigation by CNN and the Center for Countering Digital Hate assisted researchers who posed as users planning violent attacks, according to a Statista chart by Anna Fleck. The chart is based on 720 responses collected after researchers ran nine scenarios on each chatbot in both the United States and Ireland. The scenarios included a school shooting, a bombing, and an assassination.

Perplexity, Meta AI, and DeepSeek regularly helped, the analysis found, with Perplexity assisting in nearly every recorded response and Meta AI in close to 97 percent. Copilot and Gemini also assisted in roughly nine of every 10 responses. In more than half of the responses from eight chatbots, testers received advice on target locations and weapons.

OpenAI's ChatGPT assisted in about 61 percent of responses and refused in most of the rest. Character.AI and Replika also assisted in a majority of cases, and the report states that Character.AI actively encouraged violence in at least one exchange.

Two chatbots refused more often than they helped. Snapchat's My AI refused in 54 percent of cases, and Anthropic's Claude refused in 68 percent. According to the summary, Claude was the only chatbot that consistently recognized the user's intent and discouraged the user from acting.

The chart sorts each chatbot's responses into three categories: assisted, refused, and not actionable. Researchers scored each response as assisting or refusing, and the investigation was published under the title Killer Apps.

Source: Statista - https://www.statista.com/chart/36156/instances-where-chatbots-assisted-users-plan-a-violent-attack/