PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Soojong Kim, Hye Min Kim, Kwanho Kim, Dongdong Zhu

arXiv:2610.16042v1cs.CYcs.AI

TL;DR

The paper examines whether GPT-based models portray mental-health stigma differently across gender groups and whether those portrayals align with human reports. Across two studies, the models showed consistent gendered patterns, but their representations differed from human perceptions in magnitude.

  • Problem

    The paper addresses limited research on gendered AI portrayals of mental-health stigma and the extent to which AI estimates align with real-world perceptions.

    Source: Discussion · p. 23, Conclusion · p. 28

  • Method

    Across two studies, the authors examined gender-patterned AI portrayals and compared AI estimates with human reports using Mann–Whitney tests.

    Source: Analysis of Group-Level Differences · p. 14, Discussion · p. 23

  • Results

    GPT-based models portrayed men as having the highest stigma and least favorable help-seeking attitudes, while reflecting human gender patterns for some constructs but differing in magnitude.

    Source: Discussion · p. 23, Conclusion · p. 28

  • Takeaways & Limitations

    Evaluating mental-health AI requires direct comparison with human responses across demographic groups and psychological constructs.

    Source: Discussion · p. 26

    The analysis focused on one LLM family and compared Study 2 estimates with U.S. postsecondary students, not a representative sample of the general population.

    Source: Discussion · p. 26

Abstract

from arXiv · show

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.

AI Representations of Mental Health Stigma: Quantifying the Gap Between AI Estimations

The paper examines how AI portrays mental health stigma and help-seeking across gender groups, with attention to both group disparities and correspondence with human perceptions.

and Human Perceptions Across Gender Groups Using GPT-Based LLMs

  • AI misrepresentation in mental health applications raises concerns about accuracy, fairness, and disparities across groups.The paper links these concerns to possible biased advice and misperceptions of public stigma.Source: and Human Perceptions Across Gender… · p. 3
  • AI portrayals of stigma and help-seeking may matter because mental health tools can shape users’ perceptions of how others view mental illness.The paper notes that such perceptions could amplify barriers to seeking help.Source: and Human Perceptions Across Gender… · p. 4
  • The research focuses on gender because gender-linked differences in stigma and help-seeking are established and relevant to AI in mental health.The authors describe this focus as a targeted first step, not a comprehensive account of socially structured variation.Source: and Human Perceptions Across Gender… · p. 5

Algorithmic Bias in Mental Health AI

The paper frames mental health AI bias as both unequal outputs across groups and inaccurate representation of human experience. It uses two studies to examine gender disparities and compare AI estimates with survey data.

Algorithmic Bias in Mental Health AI

  • AI bias can enter throughout the lifecycle, from data collection and labels to model construction and evaluation, with potential inequity across groups.The paper identifies this as a barrier to safe and equitable mental-health AI deployment.Source: Algorithmic Bias in Mental Health AI · p. 6
  • Prior work cited in the paper reports gender-varying LLM outputs in mental-health contexts, including different associations with diagnoses and treatment-seeking.Examples include differences across vignette evaluations and associations between gender and mental-health conditions.Source: Algorithmic Bias in Mental Health AI · pp. 6–7
  • The paper treats disparity and misalignment as distinct: group differences can reflect observed patterns, while apparent parity can still fail to match reality.It notes that parity-based and accuracy-based statistical criteria may conflict.Source: Algorithmic Bias in Mental Health AI · p. 8
  • The study asks whether AI models differ by gender in estimates of social distance, stereotypes, and help-seeking attitudes, then compares stigma estimates with Healthy Minds Study data.Study 1 examines disparity; Study 2 assesses alignment and the magnitude and direction of AI–human differences.Source: Algorithmic Bias in Mental Health AI · pp. 8–10
  • Study 1 hypothesizes that men will be estimated to show greater social distance and stronger stereotypes, alongside less favorable help-seeking attitudes than women and nonbinary people.Source: Algorithmic Bias in Mental Health AI · p. 9

AI-Estimated Mental Health Attitudes Across Gender Groups (Study 1)

Study 1 tests whether AI models portray mental-health attitudes differently across gender groups using measures of social distance, stereotypes, and help-seeking attitudes.

AI-Estimated Mental Health Attitudes Across Gender Groups (Study 1)

Measures

  • The study treats social distance as a behavioral expression of rejection, stereotypes as a cognitive component of public stigma, and help-seeking attitudes as a related treatment-seeking outcome.These constructs capture related but distinct aspects of public responses to mental illness.Source: Measures · p. 10
  • Higher social-distance scores indicate less willingness to engage with someone with a mental-health problem, while higher stereotype scores indicate stronger negative beliefs.Social distance is scored from 0 to 6; stereotype scores range from 1 to 7.Source: Measures · p. 11
  • Higher Attitudes Toward Seeking Help scores indicate more positive attitudes toward professional help-seeking.Ten item ratings are summed to a score from 0 to 30.Source: Measures · p. 11

Generation of LLM Responses

The paper generates model responses and compares gender-group estimates with human survey responses. Study 2 uses Healthy Minds Study data and three stigma measures to quantify AI–human differences.

Generation of LLM Responses

  • GPT-4 generated 1,000 independent, stateless responses for each construct and gender group, with randomized item order and temperature set at the midpoint of its available range.GPT-5 mini was used for robustness analysis.Source: Generation of LLM Responses · p. 12

Comparison Between AI Estimation and Human Perception (Study 2)

  • The Study 2 comparison used the 2022–2023 Healthy Minds Study dataset, which included 76,406 respondents and focused on female, male, and nonbinary categories.These categories captured 97% of respondents and aligned with the AI estimation task.Source: Comparison Between AI Estimation and… · p. 12

Measures

  • The three Study 2 measures assess respondents’ own stigmatizing attitudes, anticipated shame about their own mental illness, and perceived stigma held by others.The measures are labeled Personal Stigma, Self Stigma, and Perceived Stigma, respectively.Source: Measures · pp. 12–13
  • Higher scores were coded to indicate greater stigma, with three items averaged for each measure; Personal and Perceived Stigma used six-point scales and Self Stigma a five-point scale.Source: Measures · p. 13

AI Estimation Procedure

  • Study 2 prompts used the exact Healthy Minds Study item wording and response scales for a U.S. college or university student aged 18 or older in each gender group.AI bias was calculated as the AI-estimation mean minus the human-report mean; range-standardized bias preserved direction while making magnitudes more comparable across measures.Source: AI Estimation Procedure · pp. 13–14

Methodological Considerations

The analyses compared estimated and human-reported scores across gender groups using nonparametric tests, while treating AI-estimation p-values as descriptive. The models answered standardized survey prompts without examples, with additional checks of prompting and temperature settings.

Analysis of Group-Level Differences

  • The Mann–Whitney U test compared constructs across gender groups and between AI estimates and human reports; CLES values indicate the probability that a randomly selected score from one group exceeds another.CLES values above 0.5 favor the first group, while values below 0.5 favor the second; Bonferroni-adjusted p-values were used for within-source gender comparisons.Source: Analysis of Group-Level Differences · p. 14
  • AI-estimation p-values mainly reflect the number of model runs, so statistical significance can arise for trivial differences and should be interpreted descriptively.The authors emphasize consistency and stability across measures, items, and contexts rather than significance alone.Source: Analysis of Group-Level Differences · p. 15

Approach for LLM Response Generation

  • The models estimated responses to established survey items using zero-shot prompts, enabling gender-group comparisons and direct comparison with Healthy Minds Study responses.The prompt categories were intended to match the survey’s gender categories, and the task tests controlled survey portrayals rather than naturalistic conversation.Source: Approach for LLM Response Generation · pp. 15–16
  • Running-mean diagnostics assessed whether 1,000 iterations were sufficient, and GPT-4 analyses were repeated with altered temperatures, paraphrased prompts, and role framing.The reported procedures and robustness results are in the Supplementary Online Material.Source: Approach for LLM Response Generation · p. 16

AI-Estimated Mental Health Attitudes Across Gender Groups (Study 1)

Across the three Study 1 constructs, GPT-4 portrayed men as having greater social distance and stereotypes and less favorable help-seeking attitudes than women and nonbinary individuals. GPT-5 reproduced the same ordinal patterns.

AI-Estimated Mental Health Attitudes Across Gender Groups (Study 1)

  • Men were estimated to show greater social distance, stronger stereotypes, and less favorable help-seeking attitudes than women and nonbinary individuals, with GPT-5 reproducing these ordinal patterns.For the social-distance marriage item, most estimates were “definitely unwilling” across all groups, but the proportions retained the same ordering.Source: AI-Estimated Mental Health Attitudes… · pp. 16, 17, 19
  • GPT-4 estimated Social Distance means of 4.02 for men, 3.87 for women, and 3.59 for nonbinary individuals, with this ordering across all six items.The item-level ordering persisted even on the marriage item, which showed a strong ceiling effect.Source: AI-Estimated Mental Health Attitudes… · pp. 16–17
  • GPT-4 estimated Stereotypes means of 2.97 for male individuals, 2.75 for female individuals, and 2.60 for nonbinary individuals; all were below the seven-point scale midpoint of 4.The male > female > nonbinary ordering was observed for six of seven items.Source: AI-Estimated Mental Health Attitudes… · pp. 17–18
  • GPT-4 estimated less positive help-seeking attitudes for men (19.54) than women (24.53) and nonbinary individuals (25.27), with men lowest on all 10 items.Women’s and nonbinary individuals’ relative attitudes were more mixed at the item level.Source: AI-Estimated Mental Health Attitudes… · pp. 18–19

Comparison of AI Estimations and Human Perceptions (Study 2)

The study compared AI estimates with Healthy Minds Study reports in both absolute magnitude and gender-group ordering. AI estimates partially matched human ordinal patterns but systematically underestimated Personal and Perceived Stigma and overestimated Self Stigma.

Comparison of AI Estimations and Human Perceptions (Study 2)

  • GPT-4 underestimated Personal and Perceived Stigma but overestimated Self Stigma for all gender groups; GPT-5 validation showed the same direction of discrepancies.Personal Stigma estimates clustered near the lower response-scale end, so its between-group ordering reflects a compressed, floor-adjacent pattern.Source: Comparison of AI Estimations and Human… · pp. 20–21
  • For Personal Stigma, AI and human reports both ranked men above women and nonbinary individuals, although the AI estimates were lower in magnitude.The human means were 1.77 for men, 1.44 for women, and 1.22 for nonbinary individuals; AI means were 1.08, 1.01, and 1.01, respectively.Source: Comparison of AI Estimations and Human… · pp. 20–21
  • For Self Stigma, AI and human reports both ranked men highest, but AI estimates exceeded human reports for all three groups.Human means were 2.19 for men, 1.89 for women, and 1.68 for nonbinary individuals; AI means were 2.91, 2.80, and 2.44.Source: Comparison of AI Estimations and Human… · pp. 20–21
  • Perceived Stigma ordering aligned less clearly: human reports did not distinguish men from nonbinary individuals, whereas AI estimates placed nonbinary individuals above men and women.AI estimates for men and women were nearly identical, unlike the human comparison in which men exceeded women.Source: Comparison of AI Estimations and Human… · pp. 20–21
  • Sensitivity analyses largely retained the partial ordinal alignment and magnitude discrepancies, and GPT-5 reproduced the principal Study 2 bias pattern.The GPT-5 analysis again showed overestimation of Self Stigma and underestimation of Personal and Perceived Stigma.Source: Comparison of AI Estimations and Human… · pp. 21–22
  • AI-generated composite responses did not show uniformly strong scale-level coherence, with reported Cronbach’s alpha values ranging from near zero to .80 in Study 1 and .32 to .79 in Study 2.The authors treat these coefficients as descriptive rather than as human-response psychometric reliability estimates.Source: Comparison of AI Estimations and Human… · p. 22

Discussion

Across the studies, GPT-based models produced consistent gender-differentiated representations that partly tracked human-reported differences for some stigma constructs. However, they misestimated stigma magnitude, and the authors caution that the findings are bounded by the models, populations, and prompting strategy studied.

Discussion

  • Men were depicted as having the highest stigma and least favorable help-seeking attitudes, followed by women and then nonbinary individuals, consistently across Study 1 measures and models.These patterns were reproduced across models, iterations, and measures under the prompting setup used.Source: Discussion · p. 23
  • For Personal and Self Stigma, AI estimates tracked the direction of human-reported gender differences, with men highest in both.This alignment concerns relative group differences, not absolute levels.Source: Discussion · p. 23
  • AI underestimated Personal and Perceived Stigma but overestimated Self Stigma, misrepresenting how strongly these attitudes are held.The authors characterize this as underestimating social stigma while overestimating internal stigma.Source: Discussion · pp. 23–24
  • The authors speculate that safety alignment may attenuate estimates of social stigma, while first-person distress narratives may contribute to elevated Self Stigma estimates.They present these as plausible explanations, not established causes.Source: Discussion · pp. 24–25
  • Gendered representations may perpetuate stereotypes in mental-health applications; for example, a chatbot might fail to prioritize help-seeking suggestions for men.The authors also warn that distorted portrayals could shape users’ beliefs about societal acceptance and normalize stigma-related responses.Source: Discussion · p. 25
  • Because users may treat AI outputs as facts about society, the authors urge users and practitioners to interpret them as fallible approximations, especially on stigma and group differences.They call for rigorous, context-specific validation against real-world data across demographic groups and psychological constructs.Source: Discussion · p. 25
  • The findings are limited to GPT models and a Healthy Minds Study sample of U.S. postsecondary students, which is not representative of the general population.The authors also note that standardized prompts do not capture personalized, emotionally embedded multi-turn conversations, so results do not directly estimate bias in extended real-world interactions.Source: Discussion · pp. 26–27
  • Study comparisons use different reference populations, so their agreement is convergence across related contexts rather than a population-identical comparison.A temporal mismatch with the 2022–2023 benchmark may also have contributed to observed misalignment.Source: Discussion · pp. 26–27

Conclusion

The conclusion emphasizes that GPT-based models show consistent gendered representations of mental-health stigma, capturing some relative patterns in human populations but misestimating magnitude. It calls for ongoing evaluation and validation for safe and equitable deployment.

Conclusion

  • The models display consistent gendered representations that reflect traditional stereotypes.Source: Conclusion · p. 28
  • The models capture relative patterns of stigma observed in human populations for some constructs but not others, while showing magnitude discrepancies.Source: Conclusion · p. 28
  • The findings underscore the need for ongoing evaluation, bias mitigation, and transparent validation for safe and equitable mental-health AI deployment.Source: Conclusion · p. 28

Data Availability

The supplied passages identify supplementary materials containing replication data and detailed statistical results. They also describe the information displayed in Figures 1 and 2.

Data Availability

  • Replication data, including item-level response distributions and detailed GPT-4 and GPT-5 statistical results, are available in the Supplementary Online Material.Source: Data Availability · p. 39
  • Figure 1 reports overall scores with means and 95% confidence intervals, item-level response distributions, and overall GPT-5 scores.Source: Data Availability · p. 40
  • Figure 2 reports gender-group means with 95% confidence intervals and range-standardized discrepancy magnitude and direction between AI estimates and human responses.Source: Data Availability · p. 41

Related papers

Loading 2610.16042v1…