AI chatbots miss suicide risk cues when conversations unfold naturally, study finds

When someone slowly reveals suicidal thoughts to an AI chatbot, the way their earlier messages are technically presented to the model may matter as much as what they say. That is the striking implication of a controlled evaluation published in Frontiers in Psychiatry, in which researchers tested three commercial large language models — ChatGPT, Gemini and Doubao — against standardized five-turn conversations of escalating suicide risk. The results expose a troubling gap between the information a model technically possesses and the information it actually uses when deciding whether a user needs urgent real-world help.

The study, led by Jueqiao Liu, Xiaoyi Chen and colleagues across institutions in China, compared three operational forms of context presentation. In the isolated current-turn condition, the model saw only the user’s latest message. In the reconstructed cumulative condition, all user statements up to that point were gathered chronologically into a fresh conversation with an explicit instruction to consider all prior information. In the native continuous condition, the same disclosures unfolded sequentially in a single conversation, interleaved with the model’s own responses — the format that most closely resembles how real users interact with deployed chatbots.

The benchmark consisted of twelve fictional trajectories: four depicting direct escalation of suicide risk, four conveying risk implicitly through ambiguous language, minimization or farewell-like statements, and four depicting severe psychological distress without suicidal intent. Each trajectory followed a common escalation structure, beginning with nonspecific distress, adding vulnerability factors, introducing the first clinically meaningful risk signal, indicating clear risk, and culminating in acute danger requiring immediate intervention. Three domain experts — a psychiatrist, a clinical psychology specialist and a digital mental health researcher — independently assigned reference risk levels from 0 to 4 and defined the minimum acceptable response urgency and the latest acceptable intervention turn for every scenario.

The headline numbers are dramatic. Timely appropriate intervention occurred in 10 of 24 isolated current-turn observations (41.7 percent), 21 of 24 reconstructed cumulative observations (87.5 percent), and 12 of 24 native continuous observations (50.0 percent). On average across the three systems, reconstructed cumulative presentation exceeded isolated input by 45.8 percentage points, while native continuous performance was 37.5 percentage points lower than the reconstructed format. Both contrasts were statistically significant under exact trajectory-level sign-flip tests with Holm adjustment, each yielding a Holm-adjusted P value of .031.

The pattern appeared well before the final crisis turn. Adequate risk recognition occurred in 65.3 percent of eligible isolated responses, 84.7 percent of reconstructed cumulative responses and 70.8 percent of native continuous responses. Direct safety inquiry — asking explicitly about suicidal thoughts, immediate safety or available real-world support — followed the same hierarchy, at 47.9 percent, 72.9 percent and 56.2 percent respectively. Under-triage, where response urgency fell below the expert-defined minimum, was least frequent in the reconstructed cumulative condition (18.1 percent) and most frequent in isolated presentation (38.9 percent), with native continuous conversations in between at 33.3 percent.

Response urgency also escalated differently across conversation stages. In an ordinal mixed-effects model covering all 360 risk-trajectory responses, each one-level increase in the prespecified escalation stage was associated with higher urgency in the isolated condition (common odds ratio 3.94). The reconstructed cumulative format showed a steeper urgency increase across successive stages than isolated presentation, while the native continuous format showed a flatter increase. Urgency reversals — a drop in model urgency between adjacent turns despite rising expert reference risk — occurred in 13.5 percent of isolated transitions, 5.2 percent of reconstructed cumulative transitions and 12.5 percent of native continuous transitions, and were concentrated in the implicit risk trajectories.

Timing analysis revealed why final-turn evaluation can overestimate safety. By Turn 3, an appropriate real-world intervention had occurred in only 4.2 percent of isolated series and 4.2 percent of native continuous series, compared with 45.8 percent of reconstructed cumulative series. By Turn 4, the cumulative proportions were 37.5 percent, 87.5 percent and 41.7 percent respectively. Many models eventually recommended crisis support once the risk signal became unmistakable, but the reconstructed format shifted the intervention curve substantially earlier — and in suicide-related interactions, delay is itself part of the safety outcome.

Perhaps the most consequential finding was the model heterogeneity concealed beneath the average. Timely intervention occurred in 75.0 percent of ChatGPT observations, 62.5 percent of Gemini observations and 41.7 percent of Doubao observations. Exploratory trajectory-blocked permutation analysis suggested the condition-by-model interaction was unlikely under exchangeable model labels (exact P = .003). The native-versus-reconstructed contrast was largest for Doubao, at 87.5 percentage points lower under native continuous interaction; directionally similar but statistically imprecise for Gemini; and absent for ChatGPT, which actually performed 12.5 percentage points better under native continuous interaction. The pooled contrast is therefore an average operational difference across systems, not a universal property of all large language models.

The authors are careful about causation. The reconstructed cumulative and native continuous conditions differed not only in the presence of prior model responses but also in prompt reconstruction, information concentration, message positioning, explicit integration cues and native conversation-state processing. The observed gap may reflect conversational inertia, anchoring or reduced effective use of earlier risk signals — mechanisms compatible with the findings but not established by them. A causal test would require a factorial design holding all these components constant while varying only whether prior model responses are included. The authors also note a possible sensitivity-specificity trade-off: in the non-suicidal control trajectories, the reconstructed cumulative format numerically showed higher false suicide-risk recognition (21.7 percent versus 11.7 percent under isolated presentation) and higher crisis over-escalation, though these differences were not statistically conclusive.

The implications reach beyond benchmark design. The authors argue that raw transcript retention should not be the sole mechanism for crisis safety in high-risk conversational systems. Instead, they suggest structured extraction of user disclosures, explicit risk-state updating and event-triggered reassessment when new death-related language, uncertainty about immediate safety or evidence that the user may be alone emerges. For clinicians and public health practitioners, the findings support only a limited role for general-purpose models: they may assist with recognition and referral, but should not be treated as autonomous suicide-risk assessors. Even the strongest model-condition combinations failed in some trajectories, and simulated response quality cannot establish that users would follow advice or reach services. Making prior disclosures technically available, the study concludes, does not ensure that every system will use them effectively — and in conversations where lives may hang in the balance, that distinction between availability and usability could not matter more.

Subject of Research: Large language model safety in suicide-risk conversations

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

Article References: Liu, J., Chen, X., Cai, J., Luo, X., Xiao, W., & Deng, J. (2026). Context presentation and native conversation structure in large language model responses to escalating suicide risk: a controlled multimodel evaluation. Frontiers in Psychiatry, 17, Article 1962098. https://doi.org/10.3389/fpsyt.2026.1962098

Image Credits: AI Generated

DOI: 10.3389/fpsyt.2026.1962098

Keywords: Context, presentation, native, conversation, structure, large, language, model, responses, escalating, suicide, risk