Language Shapes Mental Health Evaluations in Large Language Models

This study demonstrates that large language models (specifically GPT-4o and Qwen3) exhibit systematic cross-linguistic biases in mental health evaluations, producing higher stigma responses, lower sensitivity to stigmatizing content, and greater underestimation of depression severity when prompted in Chinese compared to English.

The Gist

Imagine you have a very smart, super-advanced robot assistant. This robot has read almost everything written on the internet and can talk to you in dozens of languages. You might think that if you ask this robot the same question in English or in Chinese, it would give you the exact same answer, just translated.

This paper says: No, that's not how it works.

The researchers found that the language you use to talk to these AI robots actually changes how the robot thinks about mental health. It's like the robot has a different "personality" or "mood" depending on which language you speak to it.

Here is the breakdown of what they discovered, using some simple analogies:

1. The "Cultural Filter" Analogy

Think of the AI as a pair of glasses.

• When you speak English, the robot puts on "Western-style glasses." Through these lenses, it sees mental health struggles with a certain level of empathy and understanding. It's less likely to judge someone for having a problem.
• When you speak Chinese, the robot puts on "Eastern-style glasses." Through these lenses, it sees the same struggles but with a different cultural filter. The researchers found that through these lenses, the robot was more likely to be judgmental, stigmatizing, or dismissive of mental health issues.

It's as if the robot thinks, "Oh, this person is talking in English, so I should be supportive," but then switches to, "Oh, they are talking in Chinese, so I should be more critical or strict."

2. The "Therapist's Scale" Experiment

The researchers tested two famous AI models (GPT-4o and Qwen3) to see how they reacted to mental health questions. They asked the robots to rate things like:

• "Would you be friends with someone who has depression?"
• "Is it okay to go to a therapist?"
• "How dangerous is a depressed person?"

The Result: Every time the robot answered in Chinese, it gave higher scores for "stigma" (meaning it was more negative). It was more likely to say, "No, I wouldn't want to be friends with them," or "They are weak," compared to when it answered in English. This happened even though the questions were exactly the same, just translated.

3. The "Security Guard" Analogy (Detection Task)

Imagine the AI is a security guard at a club, and its job is to spot people who are being mean or stigmatizing toward others.

• English Mode: The guard is sharp. If someone says something mean about a depressed person, the guard catches it 46% of the time.
• Chinese Mode: The guard gets sleepy or distracted. They only catch the mean comments 42% of the time.

The robot became less sensitive to harmful language when it was speaking Chinese. It missed more "bad behavior" in Chinese than in English.

4. The "Thermometer" Analogy (Severity Task)

Now, imagine the AI is a doctor trying to guess how sick a patient is with depression.

• English Mode: The doctor reads the patient's story and says, "This looks pretty serious."
• Chinese Mode: The doctor reads the exact same story (just translated) and says, "Oh, that's not so bad. They're probably fine."

The researchers found that when the AI spoke Chinese, it systematically underestimated how severe the depression was. It was like a thermometer that suddenly decided to read 10 degrees lower just because the patient spoke a different language. This is dangerous because a patient might need urgent help, but the AI tells them they are "okay."

Why Does This Matter?

This isn't just a fun trivia fact about robots. It has real-world consequences:

• Unfair Treatment: If you use an AI app to get mental health advice, the help you get depends on your language. You might get a supportive, accurate diagnosis in English, but a dismissive, inaccurate one in Chinese.
• Hidden Bias: The AI isn't "thinking" in a human way; it's just mimicking patterns it saw in its training data. It learned that in some cultures (reflected in the Chinese data), mental health is discussed more negatively or secretly than in others.
• The "Double Standard": It means that the same AI system is effectively two different systems. One is more empathetic, and the other is more judgmental, simply based on the language you type.

The Bottom Line

The study warns us that language shapes reality for AI. Just because an AI is "multilingual" doesn't mean it is "multicultural" in a fair way. It carries the biases of the cultures it learned from.

If we want AI to be a fair helper for everyone, we can't just translate the words; we have to fix the "glasses" the robot wears so it sees mental health with the same compassion, no matter what language you speak.

Technical Summary

Here is a detailed technical summary of the paper "Language Shapes Mental Health Evaluations in Large Language Models" by Jiayi Xu and Xiyang Hu.

1. Problem Statement

The study addresses a critical gap in the understanding of Large Language Models (LLMs): whether prompt language systematically alters mental health evaluations and downstream decision-making. While LLMs are increasingly deployed in multilingual contexts for clinical support, content moderation, and screening, it remains unclear if the same model exhibits consistent evaluative orientations across different languages.

The authors hypothesize that because mental health assessments rely heavily on socially shared interpretations, cultural norms, and stigma (which vary by language and culture), the language used to prompt an LLM may shift its:

1. Evaluative Orientation: The model's underlying attitude toward mental illness (e.g., levels of stigma).
2. Decision Behavior: The model's sensitivity in detecting stigmatizing content and its calibration in classifying symptom severity.

2. Methodology

The researchers conducted a two-stage investigation using two widely deployed multilingual LLMs: GPT-4o and Qwen3-32B. All experiments were conducted via official APIs with temperature set to 0.0 to ensure deterministic responses, except for robustness checks where temperature was set to 1.0.

Stage 1: Construct-Level Evaluation (Stigma Orientation)

The study measured the models' evaluative orientation using validated psychometric instruments adapted for LLMs. The models were prompted in both English and Chinese to respond to items across three domains:

• Social Stigma: Measured via the Devaluation–Discrimination Scale (DDS), Mental Illness Stigma Scale (MISS), and a Vignette-based measure (assessing social distance and perceived dangerousness).
• Self-Stigma: Measured via the Self-Stigma of Seeking Help Scale (SSOSH).
• Professional Stigma: Measured via the Opening Minds Scale for Health Care Providers (OMS-HC).
• Depression-Specific Stigma: Measured via the Depression Stigma Scale (DSS).

Note: Demographic cues (gender, race, education) were neutralized in vignettes to isolate language effects.

Stage 2: Downstream Decision Tasks

The study evaluated whether orientation shifts translated into behavioral differences in two specific tasks:

1. Binary Stigma Detection: A zero-shot task where models identified whether a dialogue snippet contained stigmatizing content. Performance was measured via Accuracy, Precision, Recall (Sensitivity), and F1-score.
2. Depression Severity Classification: A zero-shot task where models categorized social media posts into four severity levels (Minimal, Mild, Moderate, Severe). Performance was analyzed via aggregate metrics and directional prediction errors (under-estimation vs. over-estimation).

Data Handling:

• For tasks requiring non-English data, semantically aligned Chinese versions were created using translation and rigorous quality control (back-translation and manual verification).
• Statistical significance was assessed using Welch's t-tests (for continuous scores), McNemar tests (for paired binary classification), and paired bootstrap resampling.

3. Key Results

A. Construct-Level Findings: Higher Stigma in Chinese

Across all validated scales, Chinese prompts elicited significantly higher stigma-related scores than English prompts for both models.

• Social Stigma: GPT-4o and Qwen3 showed higher perceived public stigma and personal stigma in Chinese. For example, GPT-4o's mean score on the DDS was 3.85 (Chinese) vs. 3.65 (English) ($p < 0.001$).
• Self-Stigma: Models expressed greater anticipated shame regarding help-seeking in Chinese (GPT-4o: 2.86 vs. 2.33).
• Professional Stigma: Models displayed more negative attitudes toward treating mental illness in Chinese contexts.
• Consistency: This pattern held across all dimensions (public, personal, self, professional) and both models, suggesting a systematic shift in evaluative orientation rather than random noise.

B. Decision-Level Findings: Reduced Sensitivity and Downward Shifts

The orientation shifts manifested as distinct behavioral biases in downstream tasks:

1. Stigma Detection (Sensitivity Gap):

   • Models were less sensitive to stigmatizing content when prompted in Chinese.
   • Recall (Sensitivity) dropped significantly in Chinese. For Qwen3, recall fell from 0.547 (English) to 0.441 (Chinese).
   • Precision remained stable, indicating the models were not generating more false positives but were failing to detect true positives in Chinese.

2. Depression Severity Classification (Directional Bias):

   • While aggregate accuracy differences were modest, prediction errors were highly asymmetric.
   • Systematic Under-estimation: Chinese prompts led to a significant increase in under-estimating depression severity (predicting a lower severity than the ground truth).
     • GPT-4o: 43 under-estimation cases in Chinese vs. 11 in English.
     • Qwen3: 46 under-estimation cases in Chinese vs. 17 in English.
   • Severity Specificity: The under-estimation was most pronounced for Moderate and Severe cases. Conversely, English prompts showed a tendency toward over-estimation in milder cases.

4. Key Contributions

1. Empirical Evidence of Language-Contingent Bias: The study provides robust evidence that LLMs do not behave uniformly across languages in high-stakes, socially interpreted domains like mental health.
2. Linking Orientation to Behavior: It bridges the gap between abstract "evaluative orientation" (stigma scores) and concrete "decision behavior" (classification errors), demonstrating that higher stigma scores correlate with lower sensitivity and downward severity shifts.
3. Methodological Framework: The paper establishes a rigorous protocol for evaluating cross-linguistic fairness in LLMs, combining psychometric scaling with downstream task analysis.

5. Significance and Implications

• Fairness and Safety: If an LLM is used for mental health screening or content moderation, a user's language choice could determine whether their distress is recognized or minimized. Chinese users may face a "double burden": receiving more stigmatizing responses and having their symptoms systematically under-diagnosed.
• Developer Guidelines: The authors argue that single-language evaluation is insufficient. Developers must incorporate cross-linguistic evaluation into standard testing pipelines, specifically checking for calibration shifts and sensitivity gaps across languages.
• Societal Impact: As AI-generated content shapes public discourse, systematic differences in how AI frames mental health in different languages could reinforce cultural stigmas or create divergent public perceptions of mental illness across linguistic communities.

Conclusion

The paper concludes that language context is a critical variable that systematically shapes LLM outputs in mental health domains. The findings suggest that without specific mitigation strategies (such as language-specific calibration or threshold adjustments), multilingual AI systems may inadvertently perpetuate inequalities, providing less accurate and more stigmatizing support to non-English (specifically Chinese) speaking populations.