AI (statistical predictive models) work best when it's designed for a specific purpose and when the model is too challenging to derive by hand. Detecting tumors is a specific purpose, and doing so manually is challenging enough that it requires specific training. It gets a pass by me.

Predicting protein structures/drug effects: specific purpose, check. Doing it manually, yep, very challenging. Good use of AI.

LLM chatbot: purpose is unclear. Making a non-AI-based chatbot is easy and has been done before. Verdict: useless technology