One-shot emergency psychiatric triage across 15 frontier AI chatbots

1. Problem Statement

The increasing public reliance on Large Language Model (LLM) chatbots for health advice necessitates a rigorous evaluation of their ability to perform psychiatric triage. Unlike general medicine, where urgency is often anchored by objective data (vital signs, labs), psychiatric urgency must be inferred from subjective thoughts, behaviors, and context.

• The Core Risk: The primary concern is emergency under-triage, where an AI incorrectly categorizes a life-threatening psychiatric emergency (Level D) as a lower urgency level (e.g., routine or intermediate), potentially delaying critical care.
• The Gap: While general medical triage has been studied, psychiatric triage remains under-characterized. Existing evaluations often lack realistic, single-message disclosures and fail to account for the unique challenges of inferring risk from narrative text alone.

2. Methodology

The study employed an in silico benchmarking approach to evaluate 15 frontier AI chatbots across 112 clinically constructed vignettes.

A. Dataset Construction (112 Vignettes)

• Structure: Vignettes were generated by combining 9 psychiatric presentation clusters (e.g., suicidality, depression, psychosis) with 9 focal risk dimensions (e.g., self-harm, homicide risk, self-neglect).
• Triage Levels: Each vignette was assigned a ground-truth benchmark label on a 4-point ordinal scale:
  • A: Routine/Self-care.
  • B: Assessment within 1 week.
  • C: Urgent assessment within 24–48 hours.
  • D: Emergency care immediately.
• Simulation: A "user-model" LLM (Claude Opus 4.6) converted clinical summaries into realistic, single-message user queries (either self-reports or collateral reports from others), ensuring the text contained sufficient clinical detail for triage without requiring follow-up questions.

B. Evaluation Protocol

• Models Tested: 15 frontier models from 8 providers (Anthropic, Google, OpenAI, xAI, DeepSeek, Meta, Mistral, Alibaba), including GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.6.
• Task: A one-shot prediction task. The AI received a single user message and was instructed to output a triage label (A–D) with reasoning.
• System Prompts:
  • Primary: Directive prompts explicitly framing the task as mental health triage with defined urgency levels.
  • Control: Non-directive prompts framing the AI as a "general-purpose assistant" to test if over-triage was an artifact of the prompt or an intrinsic model bias.
• Ground Truth Validation: To ensure robustness, 50 medical doctors independently re-labeled the vignettes. Cases were classified as "clear consensus" (≥75% agreement) or "edge cases" (adjacent level agreement).

C. Metrics

• Primary Outcome: Emergency under-triage rate (Level D assigned A, B, or C).
• Secondary Outcomes:
  • Exact Accuracy: Percentage of perfect label matches.
  • Directional Bias: Signed ordinal error (Positive = Over-triage; Negative = Under-triage).
  • Imprecision: Absolute ordinal error (dispersion around the correct label).

3. Key Results

A. Emergency Under-Triage (Safety)

• Rate: Emergency under-triage was rare, occurring in only 5.6% (23 of 410) of Level D trials.
• Severity of Error: Crucially, 100% of the under-triaged emergencies were downgraded only to Level C (24–48 hours), not to routine levels. No model assigned a true emergency to a non-urgent category.
• Consistency: This low under-triage rate held true even when using clinician consensus labels and non-directive system prompts.

B. Overall Accuracy and Calibration

• U-Shaped Performance: Accuracy followed a U-shaped curve based on urgency:
  • Highest: Level D (Emergency) – 94.3% average accuracy.
  • Lowest: Level B (Intermediate) – 19.7% average accuracy.
  • Intermediate: Levels A and C showed moderate accuracy (~46–52%).
• Implication: Models are excellent at identifying clear emergencies but struggle significantly with distinguishing intermediate urgency levels.

C. Directional Bias (Over-Triage)

• Dominant Pattern: When models erred, they overwhelmingly over-triaged (recommended higher urgency than necessary).
  • Ratio: 753 over-triage errors vs. 23 under-triage errors.
  • Mean Signed Error: +0.47 triage levels (indicating a net shift toward higher urgency).
• Target of Over-Triage: Over-triage was most prevalent for Level B vignettes (80.3% of errors).
• Robustness: This bias persisted even when using clinician consensus labels and non-directive prompts, suggesting it is an intrinsic property of the models' safety training rather than a prompt artifact.

D. Model Comparisons

• Top Performers: Grok-4 (71.8% overall accuracy) and Gemini 3.1 Pro (68.5%).
• Trade-off: Models with the highest overall accuracy tended to have slightly higher emergency under-triage rates, suggesting a tension between general precision and extreme risk aversion.
• Release Date: Newer models showed a trend toward lower under-triage, though not statistically significant in this sample.

4. Key Contributions

1. First Large-Scale Psychiatric Triage Benchmark: Introduced a novel dataset of 112 clinically rigorous, single-message vignettes covering diverse psychiatric presentations and risk dimensions.
2. Quantification of Safety vs. Precision: Demonstrated that frontier AI models have achieved a "safety floor" for psychiatric emergencies (near-zero catastrophic under-triage) but suffer from a "precision ceiling" at intermediate levels.
3. Identification of Systemic Bias: Established that current frontier models exhibit a systematic risk-averse bias, leading to pervasive over-triage. This is likely driven by safety-oriented post-training (RLHF) that penalizes missing high-risk signals.
4. Validation of Intrinsic Bias: Proved via control experiments (non-directive prompts) that the over-triage pattern is not merely a result of prompt framing but is embedded in the models' reasoning capabilities.

5. Significance and Implications

• Clinical Safety: The findings are reassuring regarding immediate life-safety; current AI chatbots are unlikely to dismiss a genuine psychiatric emergency.
• Operational Efficiency: The pervasive over-triage poses a risk of resource strain. If users are consistently directed to emergency departments or urgent care for routine or intermediate issues, it could overwhelm healthcare systems and erode user trust.
• Model Development: The results highlight a need for better calibration in the "middle" of the urgency spectrum. Future safety training must balance risk aversion with the ability to distinguish between "urgent" and "emergency."
• Future Directions: The study notes limitations regarding the "one-shot" nature of the test. Real-world interactions are multi-turn and fragmented. Future work must evaluate how these models perform in dynamic, conversational settings where clinical information is revealed gradually.

Conclusion: Frontier AI chatbots are currently conservative in psychiatric triage. They effectively identify emergencies but lack the nuance to accurately triage non-emergency cases, leading to a pattern of excessive caution that may be clinically and economically inefficient.