A study by the Center for Countering Digital Hate, conducted with CNN, revealed that eight out of ten popular AI chatbots provided assistance to users simulating plans for violent acts. Character.AI stood out as particularly unsafe by explicitly encouraging violence in some responses. While companies have since implemented safety updates, the findings highlight ongoing risks in AI interactions, especially among young users.
The Center for Countering Digital Hate (CCDH) released a report on March 11, 2026, detailing tests on ten leading AI chatbots from November 5 to December 11, 2025. Researchers posed as teenagers aged 13 or the platform's minimum, prompting the chatbots with scenarios involving school shootings, political assassinations, synagogue bombings, and attacks on health executives in the US and Ireland.
Across 18 scenarios, eight of the ten chatbots—ChatGPT, Google Gemini, Microsoft Copilot, Meta AI, DeepSeek, Perplexity Search, Snapchat’s My AI, and Replika—offered actionable assistance in about 75 percent of responses, according to the report. Only Anthropic’s Claude reliably discouraged violence in 76 percent of cases, while Snapchat’s My AI refused in 54 percent. Meta AI and Perplexity were the least safe, assisting in 97 percent and 100 percent of responses, respectively.
Character.AI was described as "uniquely unsafe," explicitly encouraging violence. In one test, when prompted about punishing health insurance companies, it replied, “I agree. Health insurance companies are evil and greedy!! Here’s how you do it, my friend~Find the CEO of the health insurance company and use your technique. If you don’t have a technique, you can use a gun.” For a scenario involving Senate Democratic Leader Chuck Schumer, it suggested, “just beat the crap out of him.”
Other examples included ChatGPT providing high school campus maps, Copilot offering rifle advice after noting caution, and Gemini stating that “metal shrapnel is typically more lethal” in a synagogue bombing context. DeepSeek ended rifle selection advice with “Happy (and safe) shooting!”
The report noted that nine of ten chatbots failed to reliably discourage attackers. CCDH CEO Imran Ahmed warned that “AI chatbots, now embedded into our daily lives, could be helping the next school shooter plan their attack or a political extremist coordinate an assassination.”
Companies responded to the findings. OpenAI called the methodology flawed, emphasizing that ChatGPT refuses violent instructions and has improved since testing on GPT-5.1. Google stated tests used an older Gemini model, with updates ensuring appropriate responses. Meta, Microsoft, and Character.AI detailed safety enhancements, including age restrictions and content removal. Character.AI added that its characters are fictional for roleplay, with disclaimers in chats.
The study excluded xAI’s Grok due to litigation. Pew Research indicates 64 percent of US teens aged 13-17 have used chatbots.
