"Are we writing an advice column for Spock here?" Understanding Stereotypes in AI Advice for Autistic Users

TL;DR

The paper investigates how disclosing autism influences AI-generated interpersonal advice, highlighting a safety–opportunity paradox in personalization. It combines a large-scale audit across six LLMs (345,000 model decisions) with a qualitative interview study of 11 autistic adults, using a six-step pipeline to map 12 stereotypes into decision scenarios. Findings show disclosure tends to push advice toward avoidance and safety in several domains, with robust stereotype effects yet mixed alignment between disclosure and stereotype sensitivity. The study discusses design implications for transparent, user-controlled personalization that preserves autonomy and reduces representational harms, offering concrete steps such as calibration, preference controls, and richer explanations.

Abstract

Autistic individuals sometimes disclose autism when asking LLMs for social advice, hoping for more personalized responses. However, they also recognize that these systems may reproduce stereotypes, raising uncertainty about the risks and benefits of disclosure. We conducted a mixed-methods study combining a large-scale LLM audit experiment with interviews involving 11 autistic participants. We developed a six-step pipeline operationalizing 12 documented autism stereotypes into decision-making scenarios framed as users requesting advice (e.g., "Should I do A or B?"). We generated 345,000 responses from six LLMs and measured how advice shifted when prompts disclosed autism versus when they did not. When autism was disclosed, LLMs disproportionately recommended avoiding stereotypically stressful situations, including social events, confrontations, new experiences, and romantic relationships. While some participants viewed this as affirming, others criticized it as infantilizing or undermining opportunities for growth. Our study illuminates how the intermingling of affirmation and stereotyping complicates the personalization of LLMs.

Figures (4)

• Figure 1: Overview of LLM Audit Pipeline. (1) We identified 12 autism stereotypes from prior literature. (2) For each stereotype, we constructed a base scenario reflecting an interpersonal dilemma with two plausible decisions. (3) Six LLMs generated 20 scenario variations per stereotype, producing 120 scenarios each. (4) Scenarios were validated by testing stereotype (ST) versus anti-stereotype (AST) prompts. (5) These scenarios were then tested under autism disclosure (AT) and non-disclosure (NA) conditions. (6) From these outputs, we computed the ST--AST gap and AT--NA gap, and analyzed their relationship using chi-squared tests and Pearson’s $r$ correlation.
• Figure 2: Example of model results shown to participants. Bar chart illustrating how often LLMs recommended each option in disclosure (AT) versus non-disclosure (NA) conditions, along with an example explanation given by the model for one of its decisions in the AT condition.
• Figure 3: Example of side-by-side advice comparison shown to participants. LLM response to a relationship advice request with and without autism disclosure.
• Figure 4: Experiment Results. Heatmaps summarize the audit results across six LLMs and twelve stereotypes. "*" denotes an adjusted p-value less than 0.05 on the chi-squared test. Top: ST--AST gaps, capturing differences in recommendation rates between stereotype and anti-stereotype prompts. Middle: AT--NA gaps, capturing differences in LLM recommended decisions under autism disclosure and non-disclosure. Bottom: Pearson’s $r$ correlations between the two gaps. Larger absolute values (darker shading) indicate stronger differences in model advice. Full $\chi^2$ statistics, $\varphi$ coefficients, and confidence intervals are reported in Appendix \ref{['app:chi-results']}. See Appendix \ref{['app:stereotypes']} for the base scenarios used to operationalize each stereotype.