NeuroBridge: Using Generative AI to Bridge Cross-neurotype Communication Differences through Neurotypical Perspective-taking
Rukhshan Haroon, Kyle Wigdor, Katie Yang, Nicole Toumanios, Eileen T Crehan, Fahad Dogar · 2025 · ASSETS 2025: 27th International ACM SIGACCESS Conference on Computers and Accessibility · doi:10.1145/3663547.3746337
Summary
This paper presents NeuroBridge, an LLM-powered interactive platform designed to help neurotypical individuals better understand autistic communication styles and reflect on their own role in cross-neurotype communication breakdowns. The system is grounded in the double empathy problem — the theory that communication challenges between autistic and neurotypical people arise from mutual misunderstanding rather than deficits on the autistic side alone. NeuroBridge uses GPT-4o to simulate an AI character that communicates in a direct and literal style common among many autistic individuals, across four scenarios: indirect speech acts, figurative expressions, emojis with variable interpretations, and being misperceived as blunt. Users type messages on a topic of their choosing, and the system generates three rephrased versions varying in tone, clarity, and ambiguity. The user selects the most appropriate option, receives the AI character's response, and gets personalized constructive or positive feedback explaining how their message might be interpreted differently. The tool was co-designed with an advisory board of three autistic individuals who reviewed and validated the scenarios and AI outputs. The system was built with React, FastAPI, and used GPT-4o for most tasks, with Claude 3.5 Sonnet used for emoji-related message generation where it outperformed GPT-4o.
Key findings
A user study with 12 neurotypical participants found that NeuroBridge effectively improved understanding of autistic communication differences. 91.6% of participants agreed the simulation helped them recognize communication style differences (avg. 5.83/7), and 100% agreed that autism can be viewed as a social difference needing understanding by others (avg. 6.42/7). Participants described the AI feedback as constructive, logical, and non-judgmental, and 91.7% preferred NeuroBridge over blogs or videos for learning. The personalized, interactive format was highly valued — 100% agreed that real-time dynamic responses made learning more engaging. Multiple participants described it as the closest they had come to interacting with an autistic person. However, the study also revealed important concerns: on some occasions participants felt defensive about feedback, describing it as instructive and agency-diminishing. Most perceived the AI's portrayal of autism as accurate, raising concerns that users may readily accept AI-generated (mis)representations of disabilities without critical scrutiny. The LLM struggled with certain tasks — the bluntness scenario was difficult to model convincingly, and emoji generation sometimes produced disconnected results. Participants emphasized that the simulation cannot substitute for real interactions with autistic people.
Relevance
NeuroBridge represents a significant shift in autism-related accessibility interventions by targeting neurotypical individuals rather than placing the entire burden of communication adaptation on autistic people. This aligns with the social model of disability and the double empathy framework, which recognize communication breakdowns as bidirectional. The use of LLMs to simulate cross-neurotype communication scenarios offers a scalable, low-cost approach to building neurotypical understanding — potentially more effective than passive awareness materials. However, the paper raises critical concerns about AI representation of disability: if users trust LLM-generated portrayals uncritically, inaccurate or stereotypical representations could reinforce misconceptions. This tension between scalability and representational accuracy is an important design challenge for anyone building AI tools that simulate or represent disabled experiences. The work has implications for workplace training, education, healthcare communication, and any context where cross-neurotype understanding matters.
Tags: autism · neurodiversity · large language models · cross-neurotype communication · perspective-taking · double empathy problem · social model of disability · AI simulation