Research from Common Sense Media and Stanford found “systematic failures” in how chatbots recognize psychiatric conditions.