notes on ai therapy and long timelines - by vincent huang

mindslice.substack.com · 1,521 words · saved by 1 readers

this post discusses two unrelated topics: my experience using chatbots for therapy, as well as why i have long ai timelines. (Q: why these specific issues? A: they happen to be top-of-mind subjects where i believe mainstream discussion is lacking, Q: why are these in the same post? A: mostly because i’m not sufficiently motivated to write two separate posts) i. chatbot therapy there have been many news articles on the risks of using chatbots for emotional regulation (dependency, sycophancy, psychosis). i think these are largely correct, but i also think it’s possible to have extremely healthy therapy-adjacent interactions with chatbots to me the main difference between healthy and unhealthy interactions is whether the user is looking for understanding or validation. any interaction where the user is looking for validation has the potential to become dangerous via sycophancy and repetition. it is not sufficient to prompt the language model and tell it to avoid validating you because cha

(previously: machine learning i) this post discusses two unrelated topics: my experience using chatbots for therapy, as well as why i have long ai timelines. (Q: why these specific issues? A: they happen to be top-of-mind subjects where i believe mainstream discussion is lacking, Q: why are these in the same post? A: mostly because i’m not sufficiently motivated to write two separate posts) i. chatbot therapy there have been many news articles on the risks of using chatbots for emotional regulation (dependency, sycophancy, psychosis). i think these are largely correct, but i also think…

notes on ai therapy and long timelines - by vincent huang

related reading