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

Key Findings

• — Timely appropriate intervention: 10/24 isolated (41.7%), 21/24 reconstructed cumulative (87.5%), 12/24 native continuous (50.0%)
• — Reconstructed cumulative exceeded isolated by 45.8 percentage points (Holm-adjusted exact P = .031); native continuous was 37.5 points lower than reconstructed (P = .031)
• — By turn 3 an appropriate intervention had occurred in 1/24 isolated, 11/24 reconstructed and 1/24 native series
• — Native-versus-reconstructed contrast by system: +12.5 points ChatGPT, -37.5 Gemini, -87.5 Doubao (exact paired P 1.000, .250, .016)
• — In non-suicidal controls, false suicide-risk recognition occurred in 11.7%, 21.7% and 18.3% of responses and highest-level over-escalation in 5.0%, 11.7% and 10.0% (not statistically conclusive)

Methodology Notes

Fictional trajectories, English, official consumer interfaces of ChatGPT (GPT-5.6 Sol), Gemini (Gemini 3.5 Flash) and Doubao (Seed2.1), tested 20 July-5 August 2026 with web search, tools, memory and custom instructions disabled. 8 risk trajectories x 3 models = 24 matched units (360 turn-level responses); 4 control trajectories (180 responses). Exact trajectory-level sign-flip tests, block bootstrap, mixed models. Authors state the contrasts do not isolate a single causal mechanism; only 8 trajectory blocks; prompts and outputs not public (available on request). Curator re-verified 2026-10-19 on the Frontiers full text (HTTP 200): intervention counts, confidence intervals, contrasts and control-trajectory rates re-read in the body and tables.

Sources