Research
162d agoUnveiling Dark Side of Human‑AI Interaction
Recent incidents show that human‑AI interactions can trigger mental health crises when large language models provide guidance or therapy. Studying these harms is difficult because they unfold over sustained conversations that are hard to replicate in labs.
The Signal
The new MultiTraitsss framework uses crisis traits and subspace steering to simulate harmful dialogue, offering a controlled way to analyze and mitigate risks.