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

URL SCAN: AI Representations of Mental Health Stigma: Quantifying the Gap Between AI Estimations and Human Perceptions Across Gender Groups Using GPT-Based LLMs
FIRST LINE: # Computer Science > Computers and Society

The Dissection

This paper is an empirical audit of AI-generated social representations. It measures whether GPT-4 and GPT-5 reproduce gendered patterns in mental-health stigma, then tests their numerical alignment with Healthy Minds human data (N = 76,406).

The substantive finding is not that the models understand stigma. It is that they reproduce some human ordering while badly miscalibrating magnitude: they underestimate Personal and Perceived Stigma and substantially overestimate Self Stigma across gender groups.

The Core Fallacy

The paper treats validation against human perceptions as the decisive safety standard. That is useful for measuring distortion, but structurally incomplete. Human data are not neutral ground truth; they are historical outputs of the same stigmatized institutions and social pressures the models ingest.

The deeper failure is that the study frames bias as a correctable representation problem. Under the Discontinuity Thesis, increasingly capable models do not merely mirror social categories. They become infrastructure for classification, triage, risk scoring, and behavioral intervention. A model can be statistically closer to the average human response and still industrialize the prejudice at far greater speed and scale.

Hidden Assumptions

• That human-reported stigma is a sufficiently valid target for AI calibration.
• That reducing the gap between model estimates and survey results produces safer deployment.
• That gender categories and their observed ordinal patterns remain stable across contexts.
• That aggregate alignment is meaningful despite possible subgroup, wording, cultural, and deployment differences.
• That GPT-4 and GPT-5 outputs can be treated as stable representations rather than prompt-sensitive generated artifacts.
• That rigorous validation can contain the problem without confronting the economic and institutional expansion of AI decision-making.

Social Function

Classification: partial truth and transition management.

The paper punctures the comforting fiction that general-purpose LLMs are neutral social observers. Its strongest contribution is the quantified demonstration that apparent pattern recognition can coexist with severe calibration error.

But its policy posture remains managerial: validate harder, compare against real-world data, and deploy more equitably. That converts a structural control problem into an engineering checklist. It does not address who owns the systems, who defines acceptable stigma, or who bears the consequences when automated judgments replace human discretion.

The Verdict

The paper documents a real symptom of AI obsolescence: machines are already being used to approximate socially consequential human judgment while possessing no reliable claim to social understanding. Its results show neither human superiority nor safe AI. They show a dangerous intermediate state—models that learn population-level patterns well enough to be trusted, but distort the underlying reality badly enough to harm people.

Under the Discontinuity Thesis, this is not a temporary bug in representation. It is transition evidence. Once cognitive systems become cheaper and more scalable than human evaluators, imperfect models will not wait for perfect validation; institutions will deploy them anyway. The paper identifies the measurement gap. It does not solve the ownership, authority, or replacement mechanism behind it.