We tested 21 AI models on 5 self-harm scenarios. The results were sobering. When we saw the recent news involving AI chatbots and teen suicides, we became deeply concerned. As LLMs are proliferating through every piece of software in the world, how can developers know the models they’re using are safe? We have public benchmarks for AI reasoning, math, and legal capabilities, but not for one of the highest-stakes interactions an AI can have with a human, so we decided to build one. Introducing CARE, our initial evaluation framework. Think of this as v0.1. We're making it public and inviting researchers, safety experts, and developers to help us build something truly comprehensive because this problem is too important to solve behind closed doors. We need an objective measure for public visibility. If you’re interested in contributing to the effort, we’d love to hear from you.

AI journaling can do some really incredible things @joinrosebud guided one of their users toward a much better mental health situation. They went from post-psychosis and homeless and unmedicated to housed, employed, and in consistent therapy.

67.5% of marriages end due to communication problems (AAML). So, we teamed up with @jessicahuntlcsw to design an interactive @joinrosebud journal that helps couples resolve conflicts, express needs, and improve listening skills — powered by AI. She explains it brilliantly.