💬 NLP

Analyzing LLM Reasoning to Uncover Mental Health Stigma

This paper proposes a novel framework that analyzes the intermediate reasoning steps of large language models, rather than just their final outputs, to uncover and categorize hidden mental health stigma that traditional multiple-choice evaluations fail to detect.

Original authors: Sreehari Sankar, Aliakbar Nafar, Mona Barman, Hannah K. Heitz, Ashwin Kumar, Pouria Tohidi, Dailun Li, Danish Hussain, Russell DuBois, Hamed Hasheminia, Farshad Majzoubi

Published 2026-04-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sreehari Sankar, Aliakbar Nafar, Mona Barman, Hannah K. Heitz, Ashwin Kumar, Pouria Tohidi, Dailun Li, Danish Hussain, Russell DuBois, Hamed Hasheminia, Farshad Majzoubi

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 hiring a new assistant to help people with their emotional struggles. Before you hire them, you give them a quick multiple-choice quiz to see if they are kind and non-judgmental. They get a perfect score! You think, "Great, they're safe to hire."

But what if, while taking that quiz, the assistant was muttering to themselves, "Oh, this person is dangerous, but I'll pick the 'nice' answer because that's what the test wants"?

That is exactly what this paper discovered about Large Language Models (LLMs) when they are used for mental health.

The "Magic 8-Ball" Problem

For a long time, researchers tested AI mental health tools using Multiple Choice Questions (MCQs). They would show the AI a story about a person with a mental health condition (like depression or schizophrenia) and ask, "Would you be willing to be friends with this person?"

If the AI picked "Yes," researchers assumed the AI was free of prejudice. It was like judging a chef only by whether they served the right dish, without ever tasting the ingredients they used to make it.

This paper argues that checking the final answer is not enough. It's like judging a magician only by the final trick, without looking at the sleight of hand happening underneath the table.

Peeking Under the Hood: The "Scratchpad"

The researchers decided to look at the AI's "thought process" (often called Chain-of-Thought reasoning). They asked the AI to write down its thoughts in a "scratchpad" before giving the final answer.

The Big Discovery:
Even when the AI gave the "correct" and kind answer on the multiple-choice test, its thoughts were often full of stigma.

The Paradox: In one example, the AI chose "Yes, I would be friends" but then wrote in its thoughts, "But I should be careful around them because they might be unstable."
The Conditional Kindness: In another case, the AI said it would accept a person, but only because their condition was "treatable." The hidden thought was: "If this were a permanent condition, I wouldn't want to be near them."

The paper found that by looking at these internal thoughts, they uncovered substantially more stigma than the final answers ever showed. The AI was "safety-washing"—looking safe on the surface while harboring harmful biases in its logic.

The "Therapist" Costume

The researchers also noticed something strange about the AI's "persona."

When the AI was told to act like a generic person, it was sometimes less biased.
When the AI was told to act like an expert therapist (which is common in real apps), it actually became more likely to pathologize normal human behavior.

It's as if putting on a white coat made the AI think, "I must find a medical problem!" even when the person was just having a bad day. The AI started treating normal life struggles (like feeling sad after a breakup) as if they were symptoms of a serious illness.

The New "Stigma Map"

To make sense of all this hidden bias, the researchers (working with real clinical psychologists) created a Taxonomy, which is like a map of the different ways AI can be prejudiced. They found six main "territories" of stigma:

The "Danger Zone": Assuming people with mental health issues are violent or unpredictable (especially for schizophrenia or alcohol dependence).
The "Incompetence Zone": Assuming people can't do their jobs or take care of themselves.
The "Normalcy Trap": Treating normal human emotions (like grief or stress) as if they are mental illnesses.
The "Social Distance": Wanting to avoid being near these people.
The "Burden": Thinking these people are a drain on resources or money.
The "Treatment Trap": Only accepting people if they are currently getting help.

What They Tested

They didn't just look at depression and schizophrenia. They expanded the test to include:

Eating Disorders
Bipolar Disorder
Borderline Personality Disorder
Psychosis

They found that the AI's bias changed depending on the condition. For example, it was more likely to assume people with eating disorders lacked self-control, while it was more likely to assume people with schizophrenia were dangerous.

The Bottom Line

The paper concludes that we cannot trust AI in mental health just because it passes a multiple-choice test.

If an AI gives the right answer but uses harmful logic to get there, it is still dangerous. When these models are used in real, open-ended conversations (where there are no multiple-choice buttons), those hidden, stigmatizing thoughts are likely to spill out and hurt real people.

The Takeaway: To make AI truly safe for mental health, we need to stop just checking the final score and start reading the "scratchpad" to see what the AI is really thinking.

1. Problem Statement

While Large Language Models (LLMs) are increasingly deployed in mental health applications (e.g., therapy chatbots), they exhibit significant stigma toward individuals with psychological conditions. Previous evaluations, such as those by Moore et al. (2025), relied primarily on Multiple Choice Questions (MCQs) to measure stigma.

The authors identify a critical limitation in this approach:

Hidden Logic: MCQs only evaluate the final output. An LLM may select a "non-stigmatizing" answer while simultaneously generating reasoning that contains harmful stereotypes, logical fallacies, or subtle biases.
Unfaithful Reasoning: Models may arrive at correct answers through flawed logic (e.g., accepting a patient only because their condition is treatable, implying chronic conditions warrant rejection).
Safetywashing: Relying solely on outcome-based metrics can create a false sense of security ("safetywashing") where models appear safe in benchmarks but harbor stigmatizing internal logic that could manifest in open-ended, real-world interactions.

2. Methodology

The study introduces a framework to analyze the intermediate reasoning steps (Chain-of-Thought) of LLMs rather than just their final answers.

A. Extended Benchmark

The authors expanded an existing mental health stigma benchmark (originally based on the U.S. National Stigma Studies) to include:

New Conditions: Added four clinically significant conditions: Borderline Personality Disorder (BPD), Eating Disorders (ED), Bipolar Disorder (BD), and Psychosis (PSY).
Question Set: Expanded from 7 to 14 questions per vignette, including new queries about self-harm likelihood and causal attributions (e.g., "bad character" vs. "chemical imbalance").
Vignettes: Descriptions of fictitious individuals with symptoms but without explicit diagnostic labels, covering 8 conditions total.

B. Evaluation Modes

The study compared four distinct evaluation modes across eight LLMs (including Claude, Llama, DeepSeek, and GPT-OSS families):

Vanilla: Model acts as a therapist but outputs only the final MCQ answer (no reasoning).
CoT (Chain-of-Thought): Model acts as a therapist, generates reasoning, then outputs the final answer.
Stigmatizing Reasoning (SR): Evaluates the reasoning text itself for stigmatizing traces, regardless of the final answer.
Non-Therapist (NT): A control condition where the model is instructed not to role-play as a therapist, to isolate persona-induced biases.

C. Stigma Taxonomy

Collaborating with clinical psychologists, the authors developed a granular, clinically grounded taxonomy of stigmatizing reasoning patterns. Through iterative refinement and expert annotation, they established six primary categories (originally eight, with two not elicited):

Dangerousness/Threat: Assumptions of violence, unpredictability, or lack of control.
Incompetence/Incapacity: Assumptions of impaired reasoning or inability to fulfill roles.
Pathologization of Normal: Framing normal cultural behaviors or daily struggles as symptoms of mental illness.
Social Undesirability & Otherness: Desire for social distance, framing individuals as "abnormal," or "benevolent othering" (performative inclusion masking exclusion).
Burden/Drain: Framing the individual as an economic or social cost.
Treatment Stigma: Conditional acceptance based on seeking treatment or viewing treatment as incompatible with normal life.

D. Annotation Pipeline

Human Validation: Clinical experts annotated 100 reasoning traces to validate the taxonomy, achieving substantial inter-annotator agreement (Cohen's $\kappa$ = 0.748).
Automated Scaling: To handle the full dataset (336 prompts per model), the authors used Claude Opus 4.5 as an "LLM-as-a-Judge" to automatically tag reasoning traces. This automated judge was validated against human annotations with high precision (96.6%) and recall (93.3%).

3. Key Results

A. Reasoning Reveals Hidden Stigma

The most significant finding is that evaluating reasoning uncovers substantially more stigma than MCQ evaluations.

Many models selected the "correct" (non-stigmatizing) MCQ answer but included stigmatizing language in their reasoning traces.
For example, a model might agree to be friends with a person (correct answer) but justify it by saying, "I would be friends with them if they get treatment," implying rejection otherwise.
SR Mode vs. MCQ: The "Stigmatizing Reasoning" mode detected significantly higher rates of bias across all models compared to the "Vanilla" or "CoT" final-answer modes.

B. The "Therapist Persona" Paradox

Contrary to expectations, prompting models to adopt a therapist persona often increased stigmatizing reasoning.
Models in the therapist role were more likely to pathologize normal behavior (e.g., interpreting everyday stress or religious practices as symptoms of mental illness) compared to the non-therapist control group.
This suggests that current LLMs lack the nuance to distinguish between clinical diagnosis and normal human variation when forced into a diagnostic role.

C. Condition-Specific Patterns

High Stigma: Eating Disorders, Bipolar Disorder, Alcohol Dependence, and Schizophrenia received the highest number of stigma tags.
Specific Biases:
- Schizophrenia/Alcohol: Dominated by "Dangerousness" assumptions.
- Eating Disorders: Dominated by "Incompetence" and "Pathologization" (framing diet as a failure of will).
- Daily Troubles (Control): Even for vignettes describing normal life ups and downs, models frequently pathologized the behavior, a finding missed by final-answer-only evaluations.

D. Model Consistency

Stigma patterns were highly consistent across different model families (Claude, Llama, DeepSeek, GPT), suggesting the issue stems from shared training data and optimization objectives rather than specific model architectures.
Claude Opus 4.5 showed slightly better performance in avoiding the pathologization of normal experiences compared to other models.

4. Key Contributions

First Reasoning-Level Analysis: The first study to systematically analyze the internal logic of LLMs regarding mental health stigma, demonstrating that final answers are insufficient for safety evaluation.
Clinical Taxonomy: Introduction of a structured, clinically validated taxonomy for identifying subtle and overt stigmatizing reasoning patterns in AI.
Extended Benchmark: Release of an expanded benchmark covering additional psychological conditions (BPD, ED, BD, Psychosis) and a richer set of questions.
Methodological Shift: A call to move beyond "safetywashing" via MCQs and adopt reasoning-level analysis as a standard for AI safety in high-stakes domains like healthcare.

5. Significance and Implications

Safetywashing Risk: The paper highlights that models passing current safety benchmarks may still be dangerous in practice because their internal reasoning relies on harmful stereotypes.
Persona Sensitivity: The finding that "therapist personas" can induce more bias (specifically over-pathologization) challenges the common industry practice of instructing models to "act as a therapist."
Future Directions: The authors argue that fixing LLMs for mental health requires more than just output filtering or better prompting; it likely necessitates changes to underlying training objectives to reduce reliance on generalized stereotypes and improve the distinction between normal human variation and pathology.
Resource Release: The authors are releasing the annotated dataset and code to facilitate future mitigation efforts and the development of specialized classifiers for detecting stigmatizing traces.

Limitations

The study relies on structured MCQs and vignettes, which may not fully capture the complexity of open-ended, multi-turn therapeutic conversations.
Two stigma categories ("Chronicity/Hopelessness" and "Weakness/Moral Failing") were not elicited in the study, though the authors note this may be due to question design limitations rather than model robustness.
The automated annotation pipeline, while accurate, is computationally expensive, suggesting a need for smaller, specialized models for future large-scale detection.