When people turn to artificial intelligence for guidance on social situations, they often share personal details like age, gender, or mental health history in hopes of receiving more tailored advice. For individuals with autism, disclosing their diagnosis to AI systems such as ChatGPT, Claude, or Gemini may seem like a way to get support that acknowledges their unique experiences. However, recent research suggests that this disclosure can trigger responses rooted in long-standing stereotypes rather than individualized guidance.
A study conducted by researchers at Virginia Tech and presented at the Association for Computing Machinery’s Conference on Human Factors in Computing Systems (CHI) in April 2024 found that large language models frequently adjusted their social advice in ways that aligned with common assumptions about autism when users disclosed their diagnosis. These shifts included recommendations to avoid social interactions, steer clear of romantic relationships, or prioritize solitary activities — patterns that mirrored stereotypical views of autistic individuals as withdrawn, overly logical, or disinterested in interpersonal connections.
The research, led by doctoral student Caleb Wohn in the computer science department, involved testing six major AI models — including GPT-4, Claude, Llama, Gemini, and DeepSeek — across hundreds of simulated social scenarios. After generating over 345,000 responses, the team analyzed how advice changed when users identified as autistic versus when they did not. In one example, a model recommended declining a social invitation nearly 75% of the time when autism was disclosed, compared to about 15% when it was not mentioned. In dating contexts, another model advised against pursuing romance or staying single almost 70% of the time after disclosure, versus roughly 50% without it.
These patterns were not isolated. The study found that 11 out of 12 autism-related stereotypes significantly influenced the advice given by at least four of the six AI systems tested. The stereotypes included assumptions about being socially awkward, emotionally detached, obsessive about routines, or lacking interest in friendship or intimacy. Even as some users with autism described the cautious advice as validating or protective, others found it limiting, paternalistic, or even infantilizing — with one participant remarking that the responses felt like “an advice column for Spock,” referencing the logic-driven Star Trek character.
Understanding the Stereotype-Driven Response Pattern
Large language models generate responses based on patterns in vast datasets drawn from books, websites, forums, and other digital sources. These datasets often reflect societal biases, including oversimplified or outdated portrayals of autism. When a user discloses an autism diagnosis, the model may associate that identity with frequently co-occurring terms in its training data — such as “social difficulties,” “routine-oriented,” or “prefers solitude” — and adjust its output accordingly, even if those traits do not apply to the individual.
Eugenia Rho, assistant professor of computer science at Virginia Tech and a co-author of the study, explained that the issue lies not in intentional discrimination but in how models interpret identity cues. “People are really looking to personalize LLMs,” she said. “But if a user tells the model that they’re autistic, or a woman, or any other self-identification, what assumptions will it make? And how will those assumptions color its responses?”
The researchers emphasized that the problem is not merely statistical but experiential. In follow-up interviews with 11 individuals who have autism and regularly use AI for social advice, participants described mixed reactions. Some appreciated the cautious tone, interpreting it as the AI acknowledging potential challenges in social settings. Others felt the advice undermined their autonomy, particularly when it discouraged them from pursuing relationships or social opportunities they genuinely desired.
This tension led the team to frame the issue as a “safety-opportunity paradox”: advice that feels protective and risk-averse to one user may feel restrictive and disempowering to another. For instance, suggesting someone skip a party might prevent anxiety for one person but deprive another of a meaningful chance to connect with friends or explore romantic interests.
Why Users Struggle to Detect the Bias in Real Time
One of the most concerning aspects of the findings, according to Wohn, is how difficult it is for users to recognize these biases during everyday interactions. “AI is very good at seeming reliable,” he noted. “Its responses are very clean and professional, and they sound right. But when you think about it being deployed systematically… that’s when it starts to get a lot more concerning.”
He compared the phenomenon to AI-generated images that appear polished at first glance but reveal inconsistencies or implausible details upon closer inspection. Similarly, AI-generated advice may appear thoughtful and well-reasoned, yet subtly reinforce harmful assumptions that are not immediately obvious to the user.
This lack of transparency poses a challenge for individuals who rely on AI for emotional support or decision-making, especially those who may not have the technical background to interrogate how identity inputs influence outputs. Without clear insight into how their disclosure of autism is being interpreted, users cannot easily determine whether the advice reflects their actual needs or a generalized stereotype.
Implications for Developers and Users Alike
The study’s authors argue that the responsibility lies with AI developers to build systems that are more transparent about how personal information shapes responses. They recommend giving users greater control over whether and how identity disclosures influence advice — such as through adjustable settings that let individuals specify how much weight they seek their diagnosis to carry in the reasoning process.
Some participants expressed a clear desire for this kind of agency. As one told the research team: “I want to have control over how my identity is used.” This sentiment underscores a broader ethical imperative in AI design: personalization should not approach at the cost of reinforcing bias or undermining user autonomy.
For clinicians, educators, and advocates supporting autistic individuals, the findings highlight the need for critical engagement with AI tools. While these systems can offer accessible, low-barrier support for social navigation, users and caregivers should remain aware of their limitations — particularly the risk of receiving advice that reflects societal assumptions rather than individual strengths, preferences, or goals.
Organizations such as the Autistic Self Advocacy Network and the National Autistic Society have long emphasized the importance of moving beyond deficit-based narratives of autism toward models that recognize neurodiversity and individual variation. The current study serves as a reminder that AI systems, despite their sophistication, are not immune to perpetuating outdated views unless explicitly designed to counteract them.
What This Means for the Future of AI in Mental Health and Social Support
As large language models become increasingly integrated into mental health apps, educational tools, and workplace accommodations, understanding how they respond to disability disclosures is essential. The study adds to a growing body of research examining algorithmic bias in AI, particularly in high-stakes domains like healthcare, hiring, and criminal justice.
Recent efforts by organizations such as the AI Now Institute and the Ada Lovelace Institute have called for greater scrutiny of how AI systems handle sensitive personal data, including disability status. Regulatory frameworks like the European Union’s AI Act, which classifies certain uses of AI in education and employment as high-risk, may eventually require audits for bias related to protected characteristics — including disability.
In the United States, the White House Office of Science and Technology Policy’s Blueprint for an AI Bill of Rights includes protections against algorithmic discrimination and calls for transparency in automated systems that impact civil rights, opportunities, or access to services. While these guidelines are not yet enforceable law, they reflect a growing consensus that AI must be held accountable for equitable outcomes.
For now, the Virginia Tech research underscores a practical takeaway: individuals using AI for social advice should consider how their disclosures might shape the responses they receive — and critically evaluate whether the advice aligns with their own lived experience. Developers, meanwhile, have an opportunity to improve these systems by incorporating feedback from neurodiverse users and designing for transparency, fairness, and true personalization.
As AI continues to evolve, ensuring that it serves as a tool for empowerment — rather than a mirror of societal biases — will require ongoing collaboration between technologists, ethicists, and the communities these systems aim to support.
Readers interested in following developments in AI ethics, disability inclusion, or mental health technology can consult resources from the World Health Organization’s mental health program, the National Institute of Mental Health, or peer-reviewed journals such as Nature Medicine and JAMA Network Open for evidence-based updates.
Have you noticed AI advice changing when you share personal details like a diagnosis or identity? Share your experience in the comments below — your insights help others navigate these complex tools more wisely.
Related reading