Folks, there is no formula for detecting anxiety and emotions from pictures alone. Anxiety levels are correlated with nutrition and various biometric markers like cortisol. The AI, no matter how advanced can not detect cortisol levels from pictures of faces (this is before we even get to how stress and anxiety are defined and labeled).

(1) In this case they are asking people questions about pictures and having the AI score their answers, not having the AI look at a picture.

(2) There's a pretty big area of research into AI-based diagnosis, for instance, diagnosing depression based on social media posts. Given that "business as usual" in psychiatric diagnosis is terrible (conditions that affect 5% of the population get diagnosed practically 0% of the time, but everybody has autism...) it's not hard to improve matters greatly but the problem is that you can collect enough info to diagnose people without their consent.