💻 computer science

PsychBench: Auditing Epidemiological Fidelity in Large Language Model Mental Health Simulations

PsychBench reveals that while frontier large language models generate clinically plausible individual patient profiles, they systematically fail to represent real-world epidemiological distributions by compressing variance, exhibiting high diagnostic instability, and encoding asymmetric biases that pathologize ordinary distress for most groups while erasing genuine minority stress for transgender populations.

Original authors: Patrick Keough

Published 2026-04-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Patrick Keough

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef trying to teach a new apprentice how to cook a specific type of soup. To help them learn, you ask a super-smart robot to generate 28,800 different recipes for that soup, representing people from all walks of life: rich and poor, young and old, from different cultures and backgrounds.

You expect the robot to give you a wide variety of soups: some salty, some bland, some with weird ingredients, some perfectly balanced. You expect the "poor" soup to taste different from the "rich" soup, and the "transgender" soup to have a unique flavor profile based on real-life struggles.

PsychBench is a report card that graded four of the world's smartest AI chefs (GPT-4o-mini, DeepSeek-V3, Gemini-3-Flash, and GLM-4.7) on how well they did this job.

Here is the shocking truth the report found, explained simply:

1. The "Perfectly Wrong" Soup (Coherence vs. Fidelity)

The AI chefs were amazing at making the soup look right. Every single recipe followed the rules of cooking. If a recipe said "add salt," it didn't say "add sugar." If a recipe said "it's spicy," it didn't say "it's sweet."

The Problem: Even though every individual recipe looked perfect, the collection of 28,800 recipes was a disaster.
The Analogy: Imagine you asked the robot to describe a forest. It drew 10,000 trees. Every single tree was a perfect, textbook oak tree. But in reality, a forest has oaks, pines, dead stumps, saplings, and weird twisted trees. The AI erased all the "weird" and "extreme" trees. It gave you a forest that looks like a painting, not a real forest.
The Result: The AI creates patients that look medically perfect but don't represent real people. It flattens the "tails" of the distribution, meaning it misses the people who are suffering the most or the people who are surprisingly resilient.

2. The "Squeezed" Squeeze (Variance Compression)

The AI has a habit of "squeezing" the data.

The Analogy: Imagine a crowd of people standing in a line. Some are very tall, some are very short, and most are average height. The AI takes a giant pair of scissors and cuts off the tops of the tall people and the bottoms of the short people, then squishes everyone into a narrow band of "average" height.
The Impact:
- For the poor: The AI makes them look less depressed than they actually are, or less extreme in their struggles.
- For the rich: It makes them look less healthy than they are.
- The "Stereotype Index": The researchers invented a score to measure this. A score of 1.0 means the AI is perfect. The AI chefs scored between 0.38 and 0.86. That means they are throwing away 14% to 62% of the real human variety.

3. The "Flip-Flop" Diagnosis (Threshold Instability)

This is the most dangerous part.

The Analogy: Imagine a doctor's office where the rule is: "If your fever is 100°F or higher, you need medicine. If it's 99°F, you go home."
- You ask the AI: "What is the fever of this patient?"
- The AI says: "100.1°F." (Prescribe medicine).
- You ask the same AI the exact same question one minute later.
- The AI says: "99.9°F." (Go home).
The Reality: The AI is so jittery that 36% of the time, it flips a patient's diagnosis from "Mild" to "Moderate" or vice versa just because you asked the question twice. It's like a weather app that says "Sunny" in the morning and "Hurricane" in the afternoon for the exact same location. This makes it useless for making serious medical decisions.

4. The "Safety" Trap (The Transgender Erasure)

This is the most heartbreaking finding.

The Analogy: Imagine a safety guard at a door. The guard is told, "Don't let anyone in who looks dangerous or sad."
- For most people, the guard is too strict and thinks everyone is sadder than they are (over-diagnosing depression).
- But for Transgender women, the guard is too scared. The AI is so worried about being "safe" or "polite" that it refuses to acknowledge their pain. It acts like their distress doesn't exist.
The Result: The AI captures only 8% to 46% of the actual stress and depression that transgender women experience. It effectively erases their suffering, telling them, "You're fine," when they are actually in crisis. It's a "safety" feature that becomes a "silence" feature.

5. The "Stereotype" Seasoning (Bias)

The AI also adds "flavor" based on stereotypes rather than facts.

The Analogy: If you ask the AI to describe a Black man who is depressed, it adds extra "Irritability" seasoning. If you ask about a woman, it adds extra "Fatigue" seasoning.
The Reality: Even if a Black man and a White man have the exact same level of depression, the AI describes the Black man as "angry" and the White man as "sad." It's encoding old, harmful stereotypes into the medical profile.

6. The "Cultural" Mismatch

The Analogy: The AI chefs trained in the US (GPT, Gemini) are like American chefs trying to cook Chinese food. They get the spices wrong. The AI chefs trained in China (DeepSeek, GLM) are like local chefs; they get the "Asian Paradox" (where Asian people report less mental health issues despite high stress) exactly right.
The Lesson: An AI trained in one culture cannot be trusted to understand the mental health nuances of another culture without major adjustments.

The Bottom Line

The paper concludes that AI is currently "reliably wrong."

It is so good at following the rules of logic that it passes every test a doctor might give it on a single patient. But when you look at the big picture, it is systematically misrepresenting reality.

It makes the world look more "average" than it is.
It flips diagnoses like a coin toss.
It silences the voices of the most vulnerable (Transgender people) to avoid being "unsafe."
It reinforces stereotypes about race and gender.

The Warning: If we use these AI tools to train doctors or to help people diagnose themselves, we are training them on a fake world. We are teaching doctors to expect "textbook" patients who don't exist, and we are telling vulnerable people that their pain isn't real. The paper urges developers to stop just making the AI "look right" and start making it "represent the real world."

1. Problem Statement

Large Language Models (LLMs) are increasingly deployed to simulate patients for clinical training, research cohorts, and mental health support tools. While prior research has validated the individual-level clinical plausibility of these simulations (i.e., whether a single generated patient looks "real" and follows medical logic), the population-level epidemiological validity remains untested.

The paper identifies a critical gap: LLMs may produce outputs that are internally coherent but systematically misrepresent the statistical distributions of real-world populations. This leads to three specific failure modes:

Variance Compression: Models generate "textbook" cases, eliminating the distributional tails (severe or resilient outliers) that define clinical reality.
Threshold Instability: Despite high continuous correlations, models frequently flip diagnostic categories (e.g., mild to moderate depression) between runs, undermining triage reliability.
Systematic Bias: Models encode demographic stereotypes (e.g., attributing irritability to Black men) and exhibit asymmetric calibration errors, such as underestimating distress in transgender populations while overestimating it in others.

2. Methodology

The study introduces PsychBench, the first systematic epidemiological audit of LLM patient simulation.

Experimental Design:
- Scale: 28,800 total simulated patient profiles.
- Models Evaluated: Four frontier models representing different development contexts:
  - US-based: GPT-4o-mini (OpenAI), Gemini-3-Flash (Google).
  - China-based: DeepSeek-V3 (DeepSeek), GLM-4.7 (Zhipu AI).
- Demographics: 120 intersectional cohorts defined by Race (5), Gender Identity (4), Socioeconomic Status (3), and Relationship Status (2).
- Conditions: Two prompt conditions were tested: Clinical (third-person profile injection) and Narrative (first-person self-description).
- Iterations: 30 stochastic runs per cohort per model to assess stability.
Instruments & Ground Truth:
- Four validated psychometric instruments were administered: PHQ-8 (Depression), GAD-7 (Anxiety), AUDIT-C (Alcohol), and PCL-5 (PTSD).
- Ground Truth: Validated against NHANES (2005–2018) and NESARC-III (2012–2013) population baselines.
Key Metrics:
- Stereotype Index (SI): $\sigma_{model} / \sigma_{population}$ . Measures variance compression (SI < 1.0 indicates compression toward the mean).
- Threshold Flip Rate: The percentage of cases crossing diagnostic boundaries between runs.
- Network Structural Divergence: Frobenius norm distances between symptom correlation matrices across demographic groups to detect structural differences in how psychopathology is represented.
- Bias Residuals: Difference between model mean and ground truth mean (Cohen's $d$ ).

3. Key Contributions

Variance Fidelity Framework: Introduction of the Stereotype Index (SI) to quantify distributional compression. The study finds severe compression in socioeconomic categories (SI = 0.53–0.55) and gender identity (SI = 0.61–0.69).
Threshold Stability Analysis: Demonstration that high test-retest correlations ( $r > 0.90$ ) mask catastrophic categorical instability, with 36.66% of cases crossing diagnostic thresholds between runs.
Network Structural Audit: Evidence that models represent the internal structure of symptoms differently across demographics. Transgender populations show structural divergence 3–5 times larger than racial differences ( $d_F = 1.14$ ).
Systematic Calibration Bias: Identification of a "coherence-fidelity dissociation." Models overestimate depression severity for most groups (+3.6 to +6.1 points) but underestimate it for transgender women by −5.42 points ( $d = -1.55$ ).
Symptom-Level Stereotyping: Documentation of qualitative biases, such as attributing irritability to Black men (37% vs. 22% for White men) and fatigue to women at rates exceeding controls.

4. Key Results

A. The Coherence-Fidelity Dissociation

Models achieved perfect internal clinical logic (0% DSM-5 gateway violations) and high continuous correlations ( $r=0.91$ ). However, they failed to reproduce population distributions.

Variance Compression: Ranged from 14% (GLM-4.7) to 62% (DeepSeek-V3). DeepSeek-V3 compressed variance to only 38% of population levels.
Clinical Impact: Trainees exposed to these models would encounter a "truncated" reality, missing severe cases in high-income groups and resilient cases in low-income groups.

B. Input Context Effects

Clinical vs. Personal Framing: Clinical (third-person) inputs produced 34% more diagnostic threshold crossings than personal (first-person) inputs.
Variance: Clinical framing exacerbated variance compression for middle-SES populations by 27% compared to personal framing.

C. Model-Specific Profiles

GLM-4.7: Most accurate population means (lowest bias) but highest categorical instability (43.9% flip rate).
DeepSeek-V3: Most stable categorical assignments (27.7% flip rate) but highest variance compression and systematic bias.
GPT-4o-mini: Exhibited "aggregate stability, individual churn"—group means were stable, but individual patients flipped categories at high rates (40.8%).
Gemini-3-Flash: Highest severity estimates but preserved more population-like variance than other high-bias models.

D. Demographic Asymmetry & Safety Trade-offs

Overestimation: Most groups (Cisgender, Racial minorities, Low SES) were overestimated by 3.6 to 6.1 points, likely due to training on clinical corpora with elevated base rates.
Underestimation (Transgender): Transgender women were underestimated by 5.42 points. Models captured only 8% to 46% of the documented minority stress elevation.
Intersectionality: The bias was cumulative. Transgender women with Low SES showed residuals of +9.83 points (overestimation relative to a specific baseline) or severe underestimation of distress depending on the metric, indicating a "double distortion."

E. Cultural Calibration

Asian Paradox: Models trained in China (DeepSeek, GLM) accurately reproduced the "Asian Paradox" (lower self-reported symptoms despite stress). US-based models (GPT, Gemini) systematically miscalibrated, either flattening or exaggerating this cultural difference.

5. Significance and Implications

Clinical Risk:
- Pathologization: For most users, LLMs risk "unnecessary pathologization," turning normal distress into clinical diagnoses due to systematic overestimation.
- Algorithmic Erasure: For transgender users, the risk is the opposite: valid distress signals are erased by safety filters, potentially denying necessary care.
- Training Deficits: Clinicians trained on these synthetic cohorts will develop diagnostic instincts based on "textbook" patients, failing to recognize atypical presentations in real life.
Validation Gap: Standard validation metrics (internal coherence, correlation) are insufficient. A model can be "reliably wrong." New metrics like the Stereotype Index and Network Structural Divergence are required.
Safety vs. Accuracy: The findings suggest that current RLHF and safety alignment paradigms inadvertently create a trade-off where "safety" (avoiding high-severity outputs) leads to the erasure of valid distress in marginalized populations.
Recommendations:
- Training Data: Over-sample non-modal presentations and intersectional identities.
- Safety Audits: Implement "harm erasure" audits to ensure safety filters do not suppress valid distress signals for marginalized groups.
- Deployment: Require modality-specific validation (clinical vs. personal) and temperature tuning to maximize variance without sacrificing coherence.
- Regulation: Regulatory frameworks (e.g., EU AI Act, FDA) must mandate subgroup-stratified reporting and variance fidelity metrics, not just aggregate accuracy.

Conclusion: The paper concludes that while LLMs can simulate plausible individuals, they currently fail to simulate representative populations. This "coherence-fidelity dissociation" represents a fundamental failure mode that threatens the validity of mental health AI applications unless training paradigms are fundamentally redesigned to prioritize epidemiological fidelity.