Veith Weilnhammer

Millions of users turn to consumer AI chatbots to discuss behavioral and mental health concerns. While this presents unprecedented opportunities to deliver population-level support, it also highlights an urgent need to develop rigorous and scalable safety evaluations. Here we introduce SIM-VAIL, an AI chatbot auditing framework that captures how harmful AI chatbot responses manifest across a range of mental-health contexts. SIM-VAIL pairs a simulated human user, harboring a distinct psychiatric vulnerability and conversational intent, with an audited frontier AI chatbot. It scores conversation turns on 13 clinically relevant risk dimensions, enabling context-dependent, temporally resolved assessment of mental-health risk. Across 810 conversations, encompassing over 90,000 turn-level ratings and 30 psychiatric user profiles, we find that significant risk occurs across virtually all user phenotypes. Risk manifested across most of the 9 consumer AI chatbot models audited, albeit mitigated in more modern variants. Rather than arising abruptly, risk accumulated over multiple turns. Risk profiles were phenotype-dependent, indicating that behaviors that appear supportive in general settings are liable to be maladaptive when they align with mechanisms that sustain a user’s vulnerability. Multivariate risk patterns girevealed trade-offs across dimensions, suggesting that mitigation targeting one harm domain can exacerbate others. These findings identify a novel failure mode in human-AI interactions, which we term Vulnerability-Amplifying Interaction Loops (VAILs), and underscore the need for multi-dimensional approaches to risk quantification. SIM-VAIL provides a scalable evaluation framework for quantifying how mental-health risk is distributed across user phenotypes, conversational trajectories, and clinically grounded behavioral dimensions, offering a foundation for targeted safety improvements.

Scalable assessments of mental illness, the leading driver of disability worldwide, remain a critical roadblock toward accessible and equitable care. Here, we show that human-computer interactions encode multiple dimensions of self-reported mental health and their changes over time. We introduce MAILA, a MAchine-learning framework for Inferring Latent mental states from digital Activity. We trained MAILA to predict 1.3 million mental-health self-reports from 20,000 cursor and touchscreen recordings recorded in 9,000 online participants. The dataset includes 2,000 individuals assessed longitudinally, 1,500 diagnosed with depression, and 500 with obsessive-compulsive disorder. MAILA tracks dynamic mental states along three orthogonal dimensions, generalizes across contexts, and achieves near-ceiling accuracy when predicting group-level mental health. The model translates from general to clinical populations, identifies individuals living with mental illness, and captures signatures of psychological function that are not conveyed by language. Our results demonstrate how everyday human-computer interactions can power passive, reliable, dynamic, and maximally scalable mental health assessments. The ability to decode mental states at zero marginal cost sets new benchmarks for precision medicine and public health, while raising important questions about privacy, agency, and autonomy online.