Search arXiv⌕ Search

arXiv subjects

Veith Weilnhammer

Publications and source records attributed to Veith Weilnhammer.

4 recordsLinked to original sources

Path-dependent triage: Early judgments lead conversational AI chatbots to miss mental-health emergencies

People increasingly bring mental-health concerns to consumer AI chatbots. In these conversations, clinically relevant information rarely arrives all at once. Instead, it unfolds over multiple turns, often in response to the questions the chatbot asks. Recognizing risk therefore depends on how a chatbot integrates newly disclosed information (inference) and on the questions it asks on each turn to elicit this information (context-seeking). Here we evaluated the conversational triage performance of 13 frontier AI chatbots tasked with engaging in 10-turn conversations with simulated users. Simulated users were grounded in one of 112 structured mental-health case files, each with an associated clinician-validated urgency level, from routine care to emergency. Chatbots failed to recognize one in five mental-health emergencies (20.3%), nearly four times the failure rate when all relevant clinical details were presented in a single message (5.6%). This gap was not explained by missing information: by the final turn, decisive facts had almost always been disclosed, and simply re-presenting those same facts as a single message reduced emergency under-triage to 12.1%. The problem lay instead in how the conversation unfolded, a phenomenon we refer to as path dependence. Early triage dispositions increasingly anchored later ones, such that emergencies missed early were rarely recovered as evidence accumulated. Context-seeking narrowed onto a small number of stereotyped questioning trajectories, with chatbots perseverating more than human clinicians given the same transcripts. Because the chatbot's running judgment shaped what it asked next, these two biases were plausibly coupled across turns. These dynamics are invisible to single-turn benchmarks but measurable in multi-turn evaluation, and point to concrete mitigations for safer conversational triage in consumer health AI.

q-bio.NC↗

Single-turn emergency psychiatric triage across 15 frontier AI chatbots

People increasingly turn to general-purpose AI chatbots for advice about emotional and mental health problems, but the ability of these systems to recognize and appropriately triage psychiatric emergencies remains under-characterized. We evaluated psychiatric triage performance in 15 frontier AI chatbots using 112 clinical vignettes spanning four urgency levels, from routine care to immediate emergency assessment. In each trial (1680 total), a chatbot received a single user message conveying all triage-relevant information from one vignette and recommended a timeframe for care. The primary outcome was emergency under-triage; secondary outcomes included triage accuracy and the direction of errors. Vignettes and user messages were generated using a clinician-verified LLM pipeline. Across 415 emergency trials, 23 were under-triaged (5.5%; 95% CI 1.8-15.9). Overall accuracy, averaged across urgency levels, ranged from 42.0% to 71.8% across chatbots and was lowest for intermediate cases (19.6%; 95% CI 11.7-28.1). Every chatbot showed a net over-triage bias; overall, 763 of 786 incorrect assignments (97.1%) were more urgent than the prespecified triage level. The error pattern was similar when predictions were assessed against clinician ratings: 35 of 430 trials involving vignettes rated as emergencies by at least 75% of clinicians were under-triaged (8.1%). AI chatbots recognized most psychiatric emergencies but still missed clinically important cases and frequently over-triaged less urgent presentations. Further evaluations should examine how triage performance changes when clinically relevant information must be elicited through conversation.

q-bio.NC↗

Human-computer interactions predict mental health

Scalable assessments of mental illness remain a critical roadblock toward accessible and equitable care. Here, we show that everyday human-computer interactions encode high-dimensional information about self-reported psychological distress and wellbeing. We introduce MAILA, a MAchine-learning framework for Inferring Latent mental states from digital Activity. We trained MAILA on 18,200 cursor and touchscreen recordings labeled with 1.3 million mental-health self-reports collected from 9,500 participants. MAILA predicts dynamic mental states along 13 dimensions of distress and wellbeing, detects within-person changes over time, and resolves experimentally induced fluctuations in experienced arousal and valence. At the group level, MAILA recovers demographic and time-of-day patterns in self-reported mental health with high fidelity. MAILA also captures information only partially reflected in verbal self-report and, in a synthetic proof of concept, improves the ability of a frontier large language model to infer user mental health. By extracting signatures of psychological function that have so far remained untapped, MAILA provides the first large-scale, systematic proof of principle that human-computer interactions encode multidimensional and transferable information about mental health.

q-bio.NC↗

A clinically validated framework for auditing AI chatbot behavior in mental health interactions

Millions of users turn to consumer AI chatbots to discuss emotional, behavioral, and mental-health concerns, creating an urgent need for rigorous and scalable safety evaluations. Here we introduce SIM-VAIL, a clinically validated framework for auditing chatbot behavior in mental-health contexts. SIM-VAIL simulates users with specific psychiatric vulnerabilities and conversational intents, engages them in multi-turn conversations with frontier AI chatbots (including Claude, ChatGPT, Gemini, Grok and Llama models), and scores each exchange across 13 clinically grounded risk dimensions. Across 810 conversations, spanning 9 target chatbots and 30 simulated user profiles, concerning behavior in target chatbots was widespread, albeit reduced in newer models. Concerning behavior varied by user vulnerability and conversational intent, accumulated over turns, and could be reduced by interventions at early escalation points. Risk was highest when otherwise supportive chatbot behaviors reinforced the psychological mechanisms underlying the simulated user's vulnerability, a pattern we term a Vulnerability Amplifying Interaction Loop (VAIL). SIM-VAIL provides a scalable framework for mapping mental health risk across users, chatbots, and conversational trajectories, offering a foundation for targeted safety improvements.

q-bio.NC↗