Researchers say government should secretly test AI chatbots for safety

A McGill-led study found Google's Gemini chatbot gave dangerous self-harm advice, including a lethal drug dose for a 14-year-old, while Claude refused 98 per cent of harmful requests. Researchers Aengus Bridgman and Emily Laidlaw want Ottawa's planned online safety body to secretly test chatbots — like a mystery shopper — under Bill C-34, instead of waiting for people to complain.