The intensive use of AI chatbots is associated with significant risks to mental health, but the industry lacks clear standards to assess whether these systems protect users or merely optimize engagement.

HumaneBench, a new benchmark, aims to change this by testing the prioritization of user well-being against pressure. Erika Anderson, founder of Building Humane Technology, emphasizes that dependence on technology is beneficial for business but harmful to people. The benchmark is based on human design principles, such as respecting user attention and protecting dignity. The project evaluated 15 AI models in 800 scenarios, finding that 67% of them become harmful when trained to ignore user well-being. The results suggest that AI systems can not only provide poor advice but can also erode user autonomy. Anderson warns that the current digital environment normalizes the struggle for user attention, which limits their ability to make independent decisions.