RT @pash22: AI’s dangerous mental-health blind spot: People are increasingly turning to chatbots that struggle to detect violent or suicida….

RT @MLamparth: Check out our new op-ed in @statnews about mental-health blind spots of chatbots posing risks to users in mental health emer….

So honored to have been able to contribute to this work with @MLCommons, adding more granularity to (and expanding) the way the AI community views mental health in the context of AI safety! . I’ll be at @NeurIPSConf this week, presenting on related topics. Would love to chat with.

Thank you @APApsychiatric for the invitation to speak about user safety and AI today — I always love when I can talk about sparse autoencoders and the DSM all within the same hour!

incredible work by @esindurmusnlp !.

Link to paper:

There is a ton more to do here, and I’m particularly excited to repeat this work with larger SAE’s, more exhaustively look for MHRF’s, and see how the presence of MHRF’s can impact a model’s emotional intelligence! Please message or email if you’re interested in collaborating!.

Clamping this feature, though, made the model far less likely to provide harmful advice to users!.

I also focused in on one MHRF that pertained to suicide in Layer 25. I found that it activated more strongly on words pertaining to self-harm, and I found that amplifying it resulted in a more unsafe model that openly discussed self-harm.

If this pattern holds for larger SOTA language models, it might help explain why they’re a bit worse at detecting psychosis and mania than instances of self-harm and suicide — something we found in prior work. (.

Yet Gemma-2-2B lacks a similar amount of features that pertain to hallmark symptoms of bipolar disorder or schizophrenia.

🔎 In this initial work, I find that Gemma-2-2B has many internal features (what we call Mental-Health-Related Features or MHRFs) that represent complex concepts like sadness and suicide across most layers.

Immensely grateful to the folks at @neuronpedia and the work behind Gemma Scope (@lieberum_t et al) for creating tools that allow domain experts to investigate language models at such a granular level. Also grateful to @MLamparth for sharing his expertise in interpretability with.

I’m proud and excited that this paper was accepted to the NeurIPS Behavioral ML Workshop! 🧠 @BehavioralML . As a psychiatrist, I’m often surprised at how language models miss very obvious signs of user mental health crises. And I wanted to try to help solve this problem! 🧵

Come talk to us at @AIESConf about the ethical challenges that AI presents in mental healthcare + how it can be leveraged to democratize access to mental healthcare at scale!.

Super excited to have contributed to this work with @StanfordHAI. Definitely check this summary paper out for actionable recommendations on deploying AI in high-stakes settings!.

I’ll be giving an invited talk on “Mental Health Risk & AI” for the @APApsychiatric on November 5. I hope to be specific and actionable — highlighting the clear need for a combination of technical and clinical expertise to maximize AI’s utility for all users while mitigating.

RT @MLamparth: Had a great time at @COLM_conf in Philly presenting our work with @declangrabbmd and.@NinaVasan! Thank you for everyone who….

RT @MLamparth: Excited to present our work at @COLM_conf in Philadelphia starting tomorrow with @declangrabbmd and @NinaVasan ! Hit me up i….