Summary
Only 23% of AI systems designed for neurodivergent users involved neurodivergent people in the design process. This statistic explains most of the problems documented in this wiki’s AI section. When AI is designed about neurodivergent people rather than with them, the designers’ assumptions about what “normal,” “safe,” and “successful” look like become embedded in the system. Those assumptions are neurotypical.
This page collects what is known about participatory AI design with neurodivergent people: what works, what doesn’t, and what needs to change.
What the evidence shows
The 23% problem
A 2025 scoping review of AI systems designed for neurodivergent users found that the vast majority were built without meaningful neurodivergent involvement. The remaining 77% were designed about autistic people by non-autistic teams. The consequences are predictable: systems that optimise for neurotypical measures of success, that misread autistic behaviour, and that treat accommodation as an afterthought rather than a design principle.
What participatory design looks like
Systematic reviews (2023–2025) of participatory design with autistic people identify key adaptation principles:
Multiple communication modes allow text, voice, AAC, visual supports, and written communication. Relying on spoken group discussion excludes many autistic participants.
Extended timelines are necessary because short-term consultancy (a single workshop, a one-off feedback session) rarely produces meaningful input. Genuine participation requires sustained engagement over weeks or months.
Explicit power-sharing means participants have actual decision-making authority, not just advisory status. Tokenistic participation (where autistic people are consulted but their input is overridden) is worse than no participation, because it creates the appearance of consent.
Sensory-aware environments are a fundamental design requirement. Design sessions in noisy, brightly lit rooms with tight schedules exclude sensory-sensitive participants.
Compensation is essential because autistic participants contribute expertise. ASAN and other organisations emphasise that unpaid “consultation” is exploitation.
The AutSPACEs model
The AutSPACEs project (Alan Turing Institute, University of Cambridge, published May 2024 in Data & Policy) is the most thoroughly documented example of participatory AI design with autistic users. The project co-designed content moderation policies using online asynchronous discussion boards — chosen because they allow anonymity, restrict interaction to text, and accommodate varied communication needs.
The community identified anti-autistic attitudes, stigma, and prejudice as primary harm concerns — a different priority set than the platform safety teams’ typical focus on explicit content and violence. They developed a colour-coded labelling system (Green/Yellow/Red) reflecting community values rather than neurotypical safety norms. The project demonstrates that when autistic people design the rules, they produce different and — for their community — better policies.
The accommodation-normalisation distinction
The social model of disability provides the framework: AI should accommodate neurological difference, not normalise it.
Accommodation changes the environment to fit the person. An AI system that adjusts lighting and sound levels based on individual sensory profiles is accommodating. A predictive text system that learns the user’s communication style is accommodating. A schedule app that provides advance warning of changes is accommodating.
Normalisation changes the person to fit the environment. An AI system that coaches eye contact is normalising. A social skills app that rewards neurotypical behaviour is normalising. An emotion detection system that flags autistic facial expression as “abnormal” is normalising.
The same technology can do either, depending on design intent. The question to ask of any AI tool marketed for neurodivergent people is: does this help me live as I am, or does it teach me to perform as someone else?
The intellectual disability gap
Participatory design with people who have intellectual disabilities presents additional challenges. Most participatory AI research involves speaking autistic adults. Including people with minimal speech, people with severe ID, and people who communicate through behaviour rather than language requires further methodological innovation — multiple participation modes, carer-mediated input, behavioural observation, and long-term relationship-building.
This is difficult and expensive. It is also essential. The people most affected by AI systems in care settings — people with intellectual disabilities who cannot consent to monitoring, whose behaviour is interpreted by algorithms, whose daily lives are shaped by automated decisions — are the people least represented in the design process.
Principles for neurodiversity-affirming AI
Based on the evidence, AI designed for neurodivergent people should meet these criteria:
- Designed with, not about. Neurodivergent people (including people with intellectual disabilities) are involved from inception, with decision-making authority and compensation.
- Social model default. The system accommodates the person; it does not train the person to perform neurotypicality.
- Transparent. Users understand how the system works, can see what data it collects, and can opt out.
- User-controlled. The person (not the institution, not the carer, not the employer) controls the system and its data.
- Validated with the intended population. If the system is for autistic people, it must be validated with autistic people. If it is for people with intellectual disability, it must be validated with that population.
- Audited independently. Disability researchers and self-advocates (not the vendor) audit deployed systems for bias and harm.
Most existing AI systems do not meet these criteria. That is the point.
Key sources
- 2025 scoping review: 23% participation figure (Disability & Rehabilitation: Assistive Technology).
- CHI 2025: “Involvement of Autistic Adults in the Participatory Design of Technology.”
- Lowrie et al. (2024). AutSPACEs project. Data & Policy, Cambridge University Press.
- Springer (2024): “Adapting Participatory Design Activities for Autistic Adults.”