New benchmark shows AI still misses the human side of mental health care

By Hugo Francisco de SouzaReviewed by Susha Cheriyedath, M.Sc.Jul 12 2026

A new mental health benchmark reveals that today’s most advanced AI models can sound convincing, but still fall short when emotional insight, careful diagnosis, and real clinical judgment matter most.

Study: PsyEval: a comprehensive large language model evaluation benchmark for mental health. Image Credit: ahmetmapush / Shutterstock

In a recent study published in the journal npj Mental Health Research, researchers describe the development of “PsyEval,” a novel, comprehensive benchmark designed to evaluate the performance of large language models (LLMs) in the context-dependent and subjective domain of human mental health.

The benchmark was subsequently used to evaluate eleven advanced LLMs across three primary dimensions: 1. Knowledge understanding, 2. Diagnosis and assessment, and 3. Emotional support. PsyEval’s findings revealed that while some large language models can mirror human fluency and demonstrate strong factual recall, they still lack the deeper emotional insight and clinical nuance required to perform reliably across complex mental health tasks.

Specifically, PsyEval shows that the evaluated LLMs exhibit a high degree of prompt sensitivity, yet an "empathy gap" compared to human counselors. Furthermore, the findings suggest that safety-related refusals may lower the raw diagnostic classification scores of some large foundation models, whereas less guarded models may classify patients more aggressively, risking overdiagnosis.

Background

Mental illness represents one of the foremost chronic ailments in today’s human society, with data from the World Health Organization (WHO) indicating that depression alone affects 3.8% of the global population.

Compounding this issue, treatment rates remain remarkably low, at 13.7% in lower-middle-income countries, 22.0% in upper-middle-income countries, and 36.8% in high-income countries. The paper notes that these figures may be underestimated because of societal stigma and limited public awareness.

Against this background, researchers have increasingly investigated artificial intelligence (AI) and LLMs as potential tools for mental health applications. While early applications in basic sentiment analysis looked promising, evaluating these systems holistically has been challenging.

Unlike tasks with clearly verifiable outcomes, psychiatric assessment relies heavily on interpreting subtle, highly ambiguous, and often subjective verbal cues, alongside the establishment of a strong therapeutic bond between people seeking support and mental health professionals.

Unfortunately, a unified framework to stress-test whether AI can safely balance knowledge and diagnostic performance with empathetic communication was previously lacking.

About the Study

The present study aimed to address this evaluation gap by developing PsyEval, a task suite specifically designed to approximate key elements of real-world psychiatric consultations across three dimensions. The benchmark comprised separate evaluation datasets covering knowledge, diagnostic classification, and emotional support tasks.

Firstly, the ‘knowledge tasks’ dataset comprised 5,531 multiple-choice questions extracted from the United States (US) and Mainland China Medical Licensing Examinations (USMLE and MCMLE) to test core medical facts and crisis response scenarios.

Secondly, the ‘diagnostic tasks’ dataset was collated from 1,000 user-level test instances derived from selected social media posts in the Self-reported Mental Health Detections (SMHD) dataset to assess single-condition classification and multi-condition or comorbidity detection. Concurrently, 1,339 clinical dialogues from the Chinese D4 dataset were used to evaluate depression and suicide risk levels.

Finally, 1,000 user inquiries from the Chinese PsyQA platform and 939 English counseling interactions from Counsel-Chat were included as datasets for emotional support evaluation.

Subsequently, researchers applied PsyEval to compare eleven general-purpose and domain-specific LLMs on mental health-related benchmark tasks, including GPT-4, Qwen2.5-72B, and SoulChat, across three prompting strategies: 1. Formal Command, 2. Step-by-Step reasoning, and 3. Scenario Simulation.

Emotional-support responses were assessed using automatic metrics and a hybrid human-AI voting process involving human annotators and DeepSeek-V3.

Study Findings

The results showed that larger general-purpose models generally performed better on factual knowledge tasks, although performance also depended strongly on language alignment. On general knowledge tasks, Qwen2.5-72B achieved an accuracy of 91.0% on the Chinese MCMLE-mental dataset. Similarly, GPT-4-turbo led the English tasks with approximately 76.0% accuracy.

Despite this factual proficiency, model performance generally declined during urgent psychiatric scenario questions. For example, on the USMLE crisis response dataset, GPT-4-turbo's accuracy fell to approximately 73.0%.

The study further identified an "inverse scaling" phenomenon in models’ raw diagnostic classification accuracy. Smaller models like LLaMa-3-8B achieved near-perfect accuracy classifying conditions like attention-deficit/hyperactivity disorder (ADHD) at 100.0% and anxiety at 96.0%, substantially outperforming GPT-4, which achieved approximately 25% for ADHD.

The authors hypothesized that these discrepancies reflected a safety-utility trade-off between benchmark accuracy and cautious model behavior. Flagship foundation models may enforce guardrails that discourage them from diagnosing patients or practicing medicine, leading to refusals or ambiguous responses that were counted as wrong answers under PsyEval. Conversely, smaller models may have assigned diagnostic labels more readily, potentially increasing the risk of overdiagnosis.

When comparing the emotional support performance of LLMs against human counselors, the study revealed that while some leading models matched or outperformed the latter in language fluency and coherence, humans achieved an “Exploration score” of 1.85 on PsyQA and 1.93 on Counsel-Chat, consistently outperforming the evaluated AI models in asking probing questions to uncover deeper concerns.

Finally, prompt style was found to substantially influence LLM performance. Scenario Simulation prompts instructing models to adopt a mental health professional persona generally produced the highest empathy scores, whereas Step-by-Step prompts often produced more mechanical responses.

Conclusions

PsyEval demonstrates that several large LLMs achieved high accuracy on selected medical knowledge questions, although the study did not directly compare their knowledge or diagnostic performance with that of mental health professionals. The evaluated models nevertheless showed substantial differences in performance across English and Chinese tasks, remained sensitive to prompt wording, and were inconsistent when addressing comorbid conditions and grading clinical risk.

The findings from SoulChat also suggest that, in some models, highly specialized conversational fine-tuning may be associated with a loss of general knowledge performance, although this observation should not be generalized to all specialized mental health models.

As a benchmark study, the research did not prospectively test the models in patients, measure treatment outcomes, assess diagnostic safety in real-world settings, or evaluate autonomous crisis intervention.

These findings emphasize that future model development should focus on incorporating culturally diverse alignment data and designing smarter safety mechanisms that protect users without unduly reducing the models’ potential clinical utility.