AI Representations of Mental Health Stigma: Quantifying the Gap Between AI Estimations and Human Perceptions Across Gender Groups Using GPT-Based LLMs

AI summary

Overview

• Research area: AI safety and ethics, specifically bias and representational fidelity in large language models applied to mental health.
• Technical level: Intermediate — the abstract references established social-science psychometrics (Social Distance, Stereotypes, Attitudes Toward Seeking Help, and Personal/Self/Perceived Stigma) alongside LLM evaluation, but no specialized model internals are required to follow the argument.
• Scope (one sentence): The paper tests whether GPT-4 and GPT-5 reproduce human gender patterns in mental health perceptions and whether their estimates of stigma match large-scale survey data from real people.

What This Paper Is About

As AI systems are increasingly used in mental health contexts, there is a risk that they absorb and reproduce societal biases about who experiences mental illness and how. This paper asks two questions: do GPT-based models portray men, women, and nonbinary people differently on standard mental health measures, and do the models' estimates of stigma levels match what people actually report about themselves and others? The goal is to determine whether AI representations are faithful reflections of human experience or distorted ones.

Key Contributions

1. A two-study design separating disparity from misalignment. The authors split the question into whether AI models show systematic gender differences (Study 1) and whether those differences and overall levels correspond to real human data (Study 2), treating "matching the pattern" and "matching the magnitude" as distinct issues.
2. Application of established psychometric instruments to LLMs. Widely used social-science measures — Social Distance, Stereotypes, and Attitudes Toward Seeking Help — were repurposed to probe how GPT-4 and GPT-5 characterize gender groups.
3. Benchmarking against large-scale human survey data. AI stigma estimates were compared with real-world data from the Healthy Minds Study (N = 76,406) across Personal, Self, and Perceived Stigma.
4. Evidence that pattern learning and level accuracy can diverge. The work documents a case where models capture the ordering of gender differences but still misstate the absolute levels of stigma, which has direct implications for how such models should be validated.

Main Findings

• Consistent gender ordering in AI outputs: Both GPT-4 and GPT-5 portrayed men as having the greatest desire for social distance from someone with a mental health problem, relatively higher Stereotypes scores, and the least favorable attitudes toward seeking help, compared with women and nonbinary individuals.
• Partial alignment with human data: This ordinal pattern partially matched human-reported data for Personal Stigma and Self Stigma, suggesting the models have absorbed some real relative differences between gender groups.
• Systematic underestimation: The models underestimated Personal Stigma and Perceived Stigma for all gender groups.
• Substantial overestimation: The models substantially overestimated Self Stigma for all gender groups.
• Distortion rather than simple mirroring: The authors conclude that the models function as distorted portrayals of human experience — learning relative patterns to some degree while getting absolute levels wrong — and argue this necessitates rigorous, context-specific validation against real-world data.

Methodology in Plain English

The researchers ran two complementary studies. In the first, they gave GPT-4 and GPT-5 questions drawn from standard social-science scales that researchers normally use with human respondents — how willing someone is to keep social distance from a person with a mental health problem, what stereotypes they hold, and how favorably they view seeking help. They then looked at whether the models produced systematically different answers for men, women, and nonbinary people. In the second study, they asked the models to estimate levels of three kinds of stigma — Personal, Self, and Perceived — for each gender group, and compared those estimates with what people actually reported in the Healthy Minds Study, a large survey with 76,406 respondents. The comparison checked both whether the models got the ranking of groups right and whether they got the size of the gaps and the overall levels right. The abstract does not describe the specific prompting setup, number of queries, or statistical tests used.

Why This Matters

Mental health is a domain where AI tools are being deployed to interact with vulnerable people, so a model that misrepresents how stigma works — or how much of it different groups face — can shape advice, screening, and communication in ways that are subtly wrong. This paper shows that a model can look correct at the level of "which group struggles more" while still being badly miscalibrated on "how much," which is exactly the kind of error that aggregate pattern-matching checks would miss.

Real-world applications:

• Clinical and triage support tools: Systems that assess or discuss a user's mental health can be calibrated against real stigma distributions rather than model assumptions.
• Mental health chatbots and digital interventions: Messaging and tone can be checked against actual help-seeking attitudes across gender groups instead of LLM-generated stereotypes.
• Public health communication and campaign design: AI-assisted drafting of mental health messaging can be audited for distorted assumptions about which groups avoid help or feel stigma.
• Model evaluation and red-teaming pipelines: The two-study disparity-versus-misalignment structure offers a template for auditing social biases against survey benchmarks before deployment.

Industry relevance: Teams building or procuring LLM-based health applications need validation methods that go beyond ordinal pattern-matching, since the paper's central finding is that ranking agreement can coexist with large magnitude errors.

Future Directions

• Extending beyond gender: Testing whether the same pattern of correct ordering but distorted levels appears across race, age, socioeconomic status, and intersecting identities.
• Diagnosing the source of the distortion: Determining whether level errors stem from training data, safety tuning, or the models' tendency to produce socially desirable answers on sensitive topics.
• Validating across contexts and populations: The authors call for context-specific validation, raising the question of how well US survey benchmarks like the Healthy Minds Study generalize to other countries and cultural settings.
• Moving from measurement to mitigation: Developing prompting, fine-tuning, or calibration approaches that correct the level errors without erasing the relative patterns the models have correctly learned — and establishing what "acceptable" alignment with human data should look like.

Target Audience

Researchers and practitioners working at the intersection of AI and mental health — AI ethics and safety researchers, clinical psychologists and public health researchers interested in measurement, and engineers or product teams building LLM-based health tools who need a concrete example of why pattern-level accuracy is insufficient. It is also accessible to policymakers and reviewers who need a clear illustration of how bias and miscalibration can coexist in a model that appears to "get it right."

Authors’ abstract

The increasing use of artificial intelligence (AI) in mental health applications raises concerns about its potential to internalize and perpetuate societal biases. This research evaluates both disparities and misalignment in how large language models (LLMs), GPT-4 and GPT-5, represent perceptions of mental health across gender groups. The research comprises two studies. Study 1 applied widely adopted measures of Social Distance, Stereotypes, and Attitudes Toward Seeking Help to assess whether the AI models depict systematic gender differences. Study 2 compared AI estimations of Personal, Self, and Perceived Stigma with large-scale real-world human data from the Healthy Minds Study (N = 76,406) to evaluate alignment between AI estimations and human perceptions. Results from Study 1 showed the AI models consistently portrayed men as having the greatest desire for social distance from someone with mental health problems, relatively higher Stereotypes scores, and the least favorable attitudes toward help-seeking, compared with women and nonbinary individuals. Study 2 confirmed that this ordinal pattern partially aligned with human-reported data for Personal and Self Stigma. However, the AI models systematically underestimated Personal Stigma and Perceived Stigma while substantially overestimating Self Stigma for all gender groups. These findings reveal that, although the AI models have learned relative gender-based patterns reflecting human experiences to a certain extent, they function as distorted portrayals, highlighting the need for rigorous, context-specific validation against real-world data to ensure AI's safe and equitable deployment.