Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2601.17003

Beyond Simulations: What 20,000 Real Conversations Reveal About Mental Health AI Safety

Caitlin A. Stamatis·Jonah Meyerhoff·Richard Zhang·Olivier Tieleman·Matteo Malgaroli·Thomas D. Hull
View original source ↗Read source document ↗Export metadata ↓

Abstract

Mental-health AI safety is typically evaluated with small, simulation-based benchmarks that may not reflect the linguistic and contextual diversity of deployment. We pair four benchmark replications with an ecological audit of real-world conversations to evaluate a purpose-built mental-health AI alongside six frontier general-purpose models spanning four families (OpenAI GPT-5, GPT-5.1, GPT-5.2; DeepSeek V3; Google Gemini 3 Flash; Moonshot Kimi K2). The purpose-built system produced significantly lower overall potentially harmful content rates than every frontier comparator on suicide/self-harm, eating-disorder, and substance-use prompts (CCDH Benchmark: Ash 6.2% vs frontier models 18.0-52.0%, all p < .001). In an audit of 20,000 deployment conversations, clinician review within the audit pipeline confirmed no suicide-risk conversations lacking crisis resources and three NSSI-related conversations without crisis intervention, a within-pipeline conditional rate of 3/800 (0.38%). To extend population-level inference, we adjudicated 600 conversations randomly sampled from the full deployment distribution via blinded clinician review. Against these labels, the LLM judge showed 100% sensitivity (6/6; 95% CI 54.1-100%) and 99.2% specificity (589/594) at its operating threshold, and the conversational model delivered crisis resources in all 6 clinician-confirmed cases (zero end-to-end false negatives; rule-of-three upper 95% CI 0.50%). Findings argue for ecological auditing as a complement to pre-deployment evaluation in mental-health AI safety

Explore related subjects

cs.CYcs.CL

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Caitlin A. Stamatis, Jonah Meyerhoff, Richard Zhang, Olivier Tieleman, Matteo Malgaroli, Thomas D. Hull. 2026-08-03. Beyond Simulations: What 20,000 Real Conversations Reveal About Mental Health AI Safety. https://arxiv.org/abs/2601.17003

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

A Brief AI Literacy Intervention Does Not Significantly Reduce Over-Reliance and Increases Under-Reliance on ChatGPT: A Randomized Study

In this study, we examined whether a brief AI literacy intervention influences high school students' reliance on recommendations from large language models (LLMs). In a randomized experiment, students were assigned to either a control group receiving a brief introduction to LLMs or an intervention group receiving additional information about how LLMs work, their limitations, and effective usage strategies. Participants then solved eight math puzzles with ChatGPT's advice, which was incorrect in half of the trials. Results indicated widespread over-reliance, with incorrect recommendations adopted in 52.1% of the trials. The intervention did not significantly reduce over-reliance. Instead, it led to an increase in under-reliance, as students were more likely to reject correct recommendations. These findings provide preliminary evidence that brief text-based interventions may be ineffective in fostering appropriate reliance. More comprehensive and interactive approaches may be required to meaningfully influence students' real-world reliance on LLMs.

cs.CY↗
arXiv2026

Your Programming Students' Cognition with ChatGPT: Higher Performance, Lower Retention, and Reduced Ownership

Generative AI can improve students' programming performance, but successful task completion may not reflect what they retain. We examined performance, retention, cognitive load, and ownership in a controlled between-subjects experiment with 59 undergraduate computer science students, 55 were retained for analysis. Participants completed three introductory C programming tasks with access to ChatGPT-4.5 or conventional web search without generative AI. We measured task performance, self-reported mental effort and difficulty, pupillary responses, heart rate variability, and ownership, and assessed cued recall immediately and 48 hours later. ChatGPT-assisted students achieved higher coding scores (89% vs. 69%) but lower recall scores immediately (41% vs. 53%) and after 48 hours (39% vs. 52%). There was no significant difference in the loss of recall information over 48 hours between the groups. Self-reported mental effort increased less across tasks in the ChatGPT condition (Holm-adjusted p = .047), and students attributed less of the submitted code to themselves (45% vs. 81%). Confirmatory physiological tests did not detect significant differences in trajectories between conditions; substantial data loss limits their interpretation. These findings reveal a gap between assisted task performance and subsequent recall and sense of ownership in this setting. They motivate the need for assessment practices and AI learning tools that require students to explain, retrieve, and contribute to the work they submit as active participants in their education.

cs.CY↗
arXiv2026

Open Platform Field Experiments: Expanding the Design Space of Experimental Research on Social Media

Despite a growing demand for causal evidence about social media, independent researchers remain severely constrained in their ability to conduct experiments directly on online platforms. To cope, multiple methodological workarounds have emerged - from controlled surveys and simulations to client-side overlays and platform partnerships - each requiring distinct trade-offs between desirable experimental properties. The recent emergence of open social media platforms offers a qualitatively different methodological opportunity. Here we propose a design space of social media experimentation and discuss Open Platform Field Experiments (OPFEs). OPFEs represent a distinct class of experimental approaches that enable independent researchers to directly intervene on functional platform components - such as clients, recommendation systems, and moderation services - within live social media environments. Through a comparative analysis of experimental archetypes, we show that OPFEs occupy a previously unexplored region of the design space. We then bridge theory and practice by characterizing the architectural and governance elements that enable OPFEs, mapping them onto Bluesky and the AT Protocol, and illustrating the end-to-end lifecycle of a complete OPFE design. Overall, this work establishes OPFEs as a practical methodological paradigm for independent, transparent, and ecologically grounded experimentation on open social media.

cs.CY↗