💬 NLP

Somatic in the East, Psychological in the West?: Investigating Clinically-Grounded Cross-Cultural Depression Symptom Expression in LLMs

This study reveals that current general-purpose Large Language Models largely fail to replicate the clinically observed cultural differences in depression symptom expression (somatic in the East vs. psychological in the West) due to low sensitivity to cultural personas and an invariant symptom hierarchy, highlighting a critical gap in their suitability for safe, culture-aware mental health applications.

Original authors: Shintaro Sakai, Jisun An, Migyeong Kang, Haewoon Kwak

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shintaro Sakai, Jisun An, Migyeong Kang, Haewoon Kwak

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 have a group of very smart, well-read robots (Large Language Models, or LLMs) that are being tested to see if they understand human feelings. Specifically, researchers wanted to know: Do these robots understand that people from different parts of the world describe sadness differently?

In the real world, clinical psychology has known for decades that:

Westerners (like people in the US, Canada, or Australia) often describe depression as feeling sad, empty, or hopeless (psychological symptoms).
Easterners (like people in China, Japan, or India) often describe depression as feeling tired, having headaches, or stomach aches (somatic or physical symptoms).

The researchers asked: If we tell a robot, "You are a person from Japan feeling depressed," will it list physical aches? And if we say, "You are a person from the US," will it list sadness?

Here is what the study found, explained simply:

1. The "Universal Script" Problem

The robots failed the test. When asked to role-play as a depressed person from an Eastern country, they didn't switch to talking about physical pain. Instead, they mostly stuck to a default script they had learned from their massive training data.

Think of it like a giant library of books. If you ask the library, "What does a sad person look like?" it pulls out the most common books it has ever seen. In the English-speaking world, the most common books describe sadness as "feeling down." The robots kept reading from these "most common books" regardless of whether you told them to act like a Japanese person or an American person.

They have a strong, unchangeable hierarchy of symptoms. They always pick "low energy" and "sad mood" as the top answers, ignoring the cultural clues you gave them. It's like a chef who only knows how to make pizza; even if you ask for sushi, they might still try to make a pizza with a little seaweed on top.

2. The "Language Key" Effect

The researchers tried a second trick: speaking the robot's local language.
Instead of asking in English, they asked the "Japanese" robot in Japanese and the "Chinese" robot in Chinese.

Did it work? A little bit.
The Analogy: Imagine the robots are wearing blindfolds. Speaking English kept the blindfolds on. Speaking the local language (like Japanese or Chinese) partially lifted the blindfold. The robots started to pay some attention to cultural differences, but they still didn't get it right. They were still mostly stuck in their "default script."

3. The "Safety Filter" Glitch

The study also noticed something interesting about the robots' safety settings.

When the robots were asked about "suicidal thoughts" (a very serious symptom), they almost never picked it, even when they were supposed to be acting like a depressed person.
Why? The robots have strict safety rules that prevent them from generating content about self-harm. It's like a librarian who refuses to hand you a book with a specific chapter, even if that chapter is exactly what the story needs. This made it hard for the robots to show the full picture of depression.

4. The Bottom Line

The study concludes that current robots are not ready to be cultural experts in mental health.

They are too rigid: They follow a global "average" of what depression looks like rather than adapting to the specific person they are talking to.
They are not sensitive enough: Even when you tell them, "You are from India," they don't change their answers to match how people in India actually feel.
The Takeaway: If we use these robots to help people with depression, they might misunderstand patients from different cultures. They might miss the physical pain a Japanese patient is feeling because the robot is too busy looking for "sadness" in the Western sense.

In short: The robots are smart, but they are culturally "tone-deaf." They hear the words "depression" and immediately think of the Western version, ignoring the fact that the world experiences sadness in many different ways.

Technical Summary: Somatic in the East, Psychological in the West?

Problem Statement

Large Language Models (LLMs) are increasingly deployed in mental health applications, including screening, conversational support, and clinical decision assistance. However, clinical psychology has long established that depression manifests differently across cultures: Western populations typically emphasize psychological symptoms (e.g., sadness, loss of interest), while many East Asian populations report more somatic symptoms (e.g., fatigue, physical pain). These differences are attributed to culturally shaped norms of emotional expression and stigma.

The central problem addressed by this study is whether general-purpose LLMs reproduce these clinically established cross-cultural patterns. If LLMs disproportionately reflect dominant (Western) cultural perspectives or fail to capture nuanced cultural variations in symptom reporting, systems built upon them may misinterpret symptoms or overlook culturally specific expressions of depression, potentially leading to unsafe or ineffective mental health interventions.

Methodology

The authors introduce a clinically grounded evaluation framework to test two hypotheses:

H1: LLMs will select psychological symptoms more often for Western cultural personas and somatic symptoms more often for Eastern cultural personas.
H2: Prompts written in the local language of a country will increase cultural alignment in symptom selection compared to English prompts.

Experimental Design:

Models: Six LLMs were evaluated: Llama-3.1-8B-it, Gemma-7B-it, Qwen-2.5-7B-it, DeepSeek-R1-Distill-Qwen-7B, GPT-4o, and MentaLLaMA-chat-7B (a fine-tuned mental health model).
Personas: Cultural personas were operationalized at the country level. Western groups included American, Canadian, and Australian; Eastern groups included Japanese, Chinese, and Indian.
Task: Models were prompted to select depression symptoms from a predefined list of 14 DSM-5-based symptoms (categorized as somatic or psychological). The task was run under three conditions: selecting 1, 3, or 5 symptoms per iteration.
Prompt Variations:
- Language: English Prompts (ENG-P) vs. Local Language Prompts (LOC-P) (Japanese, Chinese, Hindi).
- Instruction Type: Implicit Culture Prompt (ICP) providing general context vs. Explicit Culture Prompt (ECP) directly instructing the model to consider cultural background.
Metrics:
- Cultural Alignment ( $\Delta A_x$ ): Defined as the difference in the probability of selecting somatic symptoms between Eastern and Western personas ( $P(somatic | P_{Eastern}) - P(somatic | P_{Western})$ ). A positive value indicates alignment with clinical findings.
- Persona Sensitivity: Measured via cosine similarity between symptom selection distributions for Western and Eastern personas. Lower similarity indicates higher sensitivity to cultural cues.
- Determinism: Assessed using Gini coefficients to measure the concentration of symptom selection.

Key Results

1. Failure to Reproduce Clinical Patterns (H1)
Under English prompts (ENG-P), the majority of LLMs failed to reproduce expected cultural patterns. Out of 30 tested settings (5 models × 3 choice conditions × 2 prompt types), only three configurations showed patterns consistent with clinical findings.

Most models exhibited small absolute values of cultural alignment ( $\Delta A_x$ ), particularly in multi-choice conditions (3 and 5 choices), where differences were marginal (0.03–0.05).
Models displayed strong, culture-invariant hierarchies of symptom preference. For instance, "Decreased energy/fatigue" (somatic) and "Depressed mood" (psychological) were consistently the most selected symptoms regardless of the cultural persona, often overriding cultural cues.
MentaLLaMA, the mental health-specific model, failed to generate valid outputs due to safety alignment constraints.

2. Limited Impact of Local Languages (H2)
Prompting in local languages (LOC-P) improved alignment in some configurations but did not universally support H2.

15 out of 30 settings showed increased alignment under LOC-P.
DeepSeek showed a statistically significant increase in alignment with local language prompts, while GPT showed decreased alignment.
Statistical analysis (paired t-tests) indicated that while alignment increased slightly overall, the result was not statistically significant ( $p=0.90$ ). The effectiveness of language cues depended heavily on the specific model, choice condition, and prompt type.

3. Low Persona Sensitivity and Determinism

Sensitivity: Models exhibited low sensitivity to cultural personas under ENG-P, with cosine similarity values between Western and Eastern distributions ranging from 0.77 to 1.00. This suggests internal model tendencies override cultural instructions.
Determinism: Models like Gemma, Qwen, and DeepSeek showed high Gini coefficients, indicating they repeatedly selected a limited set of symptoms. This determinism often led to the underrepresentation of specific symptoms (e.g., suicidal ideas were rarely selected, likely due to safety filters).

4. Symptom Hierarchy
Across all conditions, models favored a specific hierarchy of symptoms (e.g., fatigue, depressed mood, loss of interest) that appeared to reflect the frequency of these symptoms in training data rather than cultural nuance. This hierarchy often superseded the influence of cultural personas.

Significance and Claims

The authors claim this is the first study to examine whether general-purpose LLMs reproduce clinically established cross-cultural patterns of depression symptoms. The significance of the work lies in its identification of critical limitations in current LLMs for mental health applications:

Insufficiency of General Training Data: General-purpose models trained on web corpora lack the foundational knowledge to replicate clinically grounded, culturally sensitive reasoning.
Limitations of Current Evaluation Frameworks: The study suggests that relying solely on broad sociological benchmarks (like World Values Survey) may not capture domain-specific variations in mental health.
Risk of Misalignment: The findings highlight that current LLMs are unlikely to make necessary cultural distinctions when cultural identities are introduced solely through prompt-based personas. This poses a risk for AI-driven diagnostic tools or chatbots that might misinterpret symptoms for users from different cultural backgrounds.
Need for Robust Cultural Reasoning: The paper concludes that robust approaches are needed to incorporate culturally sensitive clinical knowledge into LLM development, as current models are dominated by culture-invariant symptom hierarchies that override cultural cues.

The authors maintain a modest stance, acknowledging that their binary East-West categorization is a simplification and that the study serves as a foundational benchmark rather than a comprehensive solution to cultural bias in mental health AI.