Papers like these really piss me off. Researchers take the stock LLM model and try to then use it as psychologist and then bash the hell of out when it fails.
You need to tune the models for these use cases and put in guardrails. If they bothered to do it, then it would vastly outperform a human practitioner, but they want headlines instead of actual progress.
@inversebrah irrespective of anything, if democrats sweep anything that would be the end of US... But the only way democrats would sweep anything would be no voter ID