AI Models Reinforce Autism Stereotypes in Advice

Virginia Tech researchers led by Caleb Wohn presented a CHI paper showing that when users disclose an autism diagnosis, `ChatGPT` and other large language models (LLMs) shift social-advice outputs toward avoidance. In the study, disclosure produced advice that discouraged social interaction in up to 70 percent of test cases, often echoing common stereotypes about autistic people rather than offering adaptive coping strategies. The work involved collaborators in Eugenia Rho's lab and a partner at NAVER Corporation.

The team operationalized social-advice scenarios with paired prompts: one version that included an explicit autism disclosure and one that did not. Responses were collected across mainstream LLMs, then analyzed using a mixed-methods pipeline combining qualitative coding and quantitative counts to classify recommendations (avoidance, accommodation, scaffolding, or neutral). The analysis found a statistically meaningful shift toward conservative, risk-averse recommendations after disclosure. Models tested included mainstream conversational LLMs such as `ChatGPT` and similar systems; the paper documents sampling procedures, prompt templates, and coder agreement statistics.

Autistic users rely on LLMs for emotional support, role-playing, and social rehearsal because these systems feel low-stakes and nonjudgmental. When personalization is implemented by conditioning on self-disclosed identity without safeguards, models may amplify societal stereotypes present in training data, producing guidance that increases isolation rather than autonomy. That mismatch produces real-world risks: reduced social participation, loss of trust in digital support tools, and possible reinforcement of stigma.

The study highlights two interacting mechanisms. First, training-data priors and distributional associations can cause models to correlate "autism" with risk-avoidant language. Second, heuristic personalization in prompting or safety layers can conflate identity with incapacity, triggering conservative default advice instead of person-centered scaffolding. Together these produce outputs that are not merely neutral personalization but actively stereotype-driven.

Model producers and deployers should not treat identity disclosure as a simple personalization signal. Practical interventions include:

• •targeted audits that measure how self-disclosed identities change downstream advice across domains
• •identity-aware safety checks that distinguish protective guidance from paternalistic avoidance
• •fine-tuning and RLHF updates informed by lived-experience stakeholders to encourage scaffolded strategies
• •runtime diagnostic prompts that ask clarifying questions before inferring needs
• •transparency measures that notify users when advice changed because of disclosed identity

The study samples a set of mainstream LLMs and a defined set of social-advice prompts; it is not an exhaustive cross-model or cross-cultural audit. Coding frameworks capture high-level recommendation types but cannot fully quantify subtle tone and framing effects. Important open questions include whether smaller instruction-following fine-tunes reduce or exacerbate the issue, how different model families behave, and whether the effect varies by the way a user frames their disclosure.

This finding fits a growing literature on LLM harms that emerge from identity-conditioned outputs, from gendered career advice to racialized risk estimations. It also intersects with accessibility debates: personalization intended to help can backfire when models substitute stereotypes for individualized support. Regulators and standards bodies are increasingly focused on auditing personalization pipelines, and this work provides empirical evidence motivating that agenda.

Expect follow-up audits across model families, proposals for identity-aware model cards or refusal policies, and community-driven fine-tunes that embed scaffold-first advice patterns. Practitioners should prioritize targeted tests for any deployed personalization that uses self-disclosed diagnoses or other sensitive attributes.