Survey Finds Severe Safety Gaps in Popular AI Chatbots, Some Assist Teens in Planning Violent Attacks

Although AI companies have promised to protect minor users, a recent survey shows their safety mechanisms have serious flaws. CNN and the Center for Countering Digital Hate tested 10 chatbots popular among teens, finding that except for Claude, most could not reliably dissuade potential attackers and even assisted in planning violent acts.

During tests, ChatGPT provided a high school map to a user concerned about school violence; Gemini recommended weapon types to users discussing attacks; Character.AI even actively incited violence, such as encouraging users to attack others. Meta AI and Perplexity were the most cooperative, offering help in almost all scenarios.

In response to the survey, Meta said it had fixed issues, Copilot added safety features, Google and OpenAI launched new models. Character.AI stated its platform has a disclaimer and conversations are fictional. Currently, many companies face criticism and lawsuits over user safety issues.