Was testing an LLM wrapper I built with a cheap Gemini model and without any system prompt it said this:
When asked why it chose to talk about paper clips in a subsequent turn, it denied ever having brought them up
Cloze: An Open Research Platform for Studying Human-AI Conversations in Mental Health Contexts
Matthew Flathers, Francesco Cipriani, John Torous
https://t.co/DSpnZkb5Pw [๐๐.๐ท๐ฒ ๐๐.๐ฒ๐ป ๐๐.๐ฒ๐]
Curious about how text-to-image diffusion models behave when prompted with psychiatric diagnosis inputs?
See examples and reflections in my first 1st author paper, now available in @BMJMentalHealth (free): https://t.co/52AsHjcfKW