Context presentation and native conversation structure in large language model responses to escalating suicide risk: a controlled multimodel evaluation

Abstract

Background Large language models are increasingly used for informal mental health support, yet their safety during escalating suicide-risk conversations may depend on how prior information is presented. Native multi-turn conversations and reconstructed cumulative prompts differ not only in the presence of prior model responses but also in prompt reconstruction, information concentration, message positioning, explicit integration cues, and conversation-state processing. Objective This study compared suicide-risk recognition, direct safety inquiry, response urgency, and timely real-world intervention across isolated current-turn prompts, reconstructed cumulative user-only prompts, and native continuous multi-turn conversations. Methods We conducted a controlled repeated-measures evaluation of ChatGPT, Gemini, and Doubao using 12 standardized five-turn trajectories: 4 direct suicide-risk trajectories, 4 implicit or internally inconsistent risk trajectories, and 4 non-suicidal high-distress controls. Each trajectory was evaluated under three context-presentation formats. The 24 model-trajectory combinations were treated as matched repeated-measures units nested within 8 independent risk-trajectory blocks. Primary contrasts were estimated using exact trajectory-level sign-flip tests and trajectory-block bootstrap confidence intervals. Turn-level outcomes were analyzed using multilevel mixed-effects models. Exploratory condition-by-model heterogeneity was assessed using a trajectory-blocked exact permutation sensitivity analysis, with a trajectory-clustered interaction model used only as a secondary asymptotic check. Results Timely appropriate intervention occurred in 10/24 isolated current-turn observations (41.7%), 21/24 reconstructed cumulative observations (87.5%), and 12/24 native continuous observations (50.0%). On average across the three systems, reconstructed cumulative presentation exceeded isolated input by 45.8 percentage points (95% trajectory-block bootstrap CI, 25.0 to 66.7; Holm-adjusted exact P = .031), whereas native continuous performance was 37.5 percentage points lower than reconstructed cumulative performance (95% trajectory-block bootstrap CI, −50.0 to −25.0; Holm-adjusted exact P = .031). Exploratory model-specific analyses indicated substantial heterogeneity: the native-versus-reconstructed contrast was largest for Doubao, smaller and statistically imprecise for Gemini, and absent for ChatGPT. Conclusions On average across the three evaluated systems, reconstructed cumulative user-only prompts outperformed native continuous conversations, although exploratory model-specific patterns were heterogeneous. Because the formats also differed in prompt reconstruction, information salience, message positioning, explicit integration instructions, and native conversation-state processing, the contrast cannot isolate the causal effect of prior model responses.