How AI Bias Can Exclude Neurodivergent People
Bias can hide inside a useful tool
AI can help people write, search, summarize and find a way into a task that once felt blocked. I use these tools, and I know why a person who has spent years adapting to unfriendly systems might value them. The useful part does not remove the need to ask whose experience the system has learned to recognise.
A model may seem neutral because its answer appears instantly and confidently. Its training data, labels, evaluation measures and deployment choices all come from people and institutions. The result can reproduce old assumptions or create new ways to exclude someone. This matters when an AI tool is used to screen applicants, interpret communication, recommend support, or decide what content deserves attention.
Neurodivergent people do not form one uniform group. An autistic person, someone with ADHD, a dyslexic reader, and a person with dyspraxia may each use a system differently. Even two people with the same diagnosis may want different forms of help. Good design begins by asking what the tool is for, what harm an error could do, and who is missing from the test.
What does AI bias mean here?
Bias is not limited to an openly insulting output. It can be a system that treats one communication style as the normal one. It can rank polished, conventional wording above clear but unusual wording. It can mistake a pause, a literal answer, or an alternative route through a task for a lack of skill. It can offer a generic recommendation that ignores a person’s context.
These are possible failure modes, not proof that every product makes those errors. The outcome depends on the system, data, setting and decisions made around it. We need to examine a particular tool before claiming that it discriminates. NIST’s AI Risk Management Framework describes trustworthy AI in terms that include validity, transparency, accountability and the management of harmful bias. Its bias publication also treats bias as a problem across the lifecycle, not one bad data set that can be cleaned once.
An AI response can be factually wrong, stereotyped, inaccessible or unsuitable for an important decision. Those are different problems, and each needs its own check. “The algorithm said so” is never an explanation of what evidence a person should trust.
The idea of a normal user gets built into data
A product team may train a model on previous examples of successful applications, typical customer conversations, or standard ways of asking a question. If the original process favoured people who wrote and behaved in a narrow way, its history is a poor definition of merit. A model can learn the pattern without anyone explicitly programming a rule about neurodivergence.
Labels are another choice. Who decided that an answer was “professional,” “engaged,” or “relevant”? A short, direct response may be appropriate, but a reviewer might prefer a long social preamble. If that preference becomes a label for training or evaluation, the system can reproduce it. Proxies can matter too: a feature may correlate with a disability or with access to support without naming either one.
There is no simple fix called “more data.” More of the same records may reinforce the same pattern. Teams need to know what the data represents, where it came from, what consent and privacy limits apply, and whose experience is absent. In high-stakes settings, the question is whether the model should make that decision at all.
A fluent answer can still misread a person
Generative AI often produces confident prose. That fluency can make an uncertain interpretation feel settled. If someone asks for support after a difficult interaction, a chatbot might offer a tidy explanation of what another person “really meant.” It has not witnessed the event and cannot know the other person’s intention. A user can take that apparent certainty into a real relationship or workplace.
A person who uses a direct style may also be told to make every message warmer, more conventional, or less “intense.” Sometimes editing for an audience is useful. But a tool that always treats difference as a defect can turn assistance into pressure to mask. Ask whether the suggestion clarifies the message or simply hides the person who wrote it.
The same caution applies to summaries. An AI summary of a long conversation may omit caveats, priorities or the speaker’s uncertainty. If the summary becomes the only record someone reads, the omission can shape a decision. Keep the original available, invite correction, and make important decisions reviewable.
Accessibility is part of fairness
A system can return a fair answer and still be difficult to use. Imagine a chat interface that times out before a person finishes, a form that loses a draft, or a response that appears only as a long unstructured block. A person may never get to the promised benefit.
Clear headings, predictable controls, good keyboard support, readable text, and a way to revisit the answer all matter. A transcript or text route can help someone who cannot use audio; audio may help someone who struggles with dense text. The W3C’s guidance for cognitive and learning disabilities asks designers to involve users and make steps clear. WCAG 2.2 supplies testable accessibility criteria alongside that usability work.
Do not infer what format a person wants from a diagnosis. Offer choices where they genuinely help and remember the user’s preferences where appropriate. The service should carry some of the work, rather than asking the individual to adapt to its defaults every time.
Where the stakes change the review
Using AI to brainstorm a headline is different from using it to decide who gets an interview, what support a student receives, or whether someone’s health concern is urgent. The more consequential the outcome, the stronger the need for evidence, privacy protection, human accountability and a meaningful way to challenge an error.
A human in the loop is only useful if that person has time, authority and enough information to disagree with the model. A rubber stamp does not make a decision fair. If the tool produces a score, ask what it measures, whether it is valid for this use, and how performance varies across relevant groups. Do not use disability disclosure as the only route to discovering a problem; people may not want to share a diagnosis.
NIST’s framework recommends mapping context, measuring risks, governing the system and managing harms. Those verbs point to ongoing work. A tool that behaved acceptably in a small test can change as the model, users or workflow change. Review after deployment and provide a route for complaints and correction.
What a team can actually test
Start with a real task and write down the intended benefit. Then list plausible failures: a wrong fact, a stereotyped suggestion, an inaccessible control, loss of context, or a recommendation presented with too much certainty. Choose test cases that reflect varied communication styles and ways of completing the task. Include people with relevant lived experience and pay them for their time.
Compare outcomes, but do not reduce a person to a diagnosis or one number. Look at whether users can understand the recommendation, correct it, and reach a person when necessary. Test with the full interface, not just model outputs copied into a spreadsheet. A good response is of little use when the person cannot navigate to it.
Keep records of failures and changes. Make it possible to turn off a feature that is causing harm. When a vendor supplies the model, ask for its documentation, limitations and evaluation evidence. An opaque vendor claim of fairness is not a substitute for testing the product in your setting.
What a person using AI can do
For everyday drafting, treat a suggestion as a draft. Check names, dates, citations and claims against the original source. If a response offers a judgment about you or someone else, ask what it is based on and whether a different explanation fits. You do not owe a tool your diagnosis or private history to get help with wording.
Try giving a bounded instruction: “Keep my direct style, fix the unclear sentence, and tell me what you changed.” That leaves more of your voice intact than “make this sound professional.” Save the original. If a tool repeatedly pushes you toward a version of yourself you do not recognize, you can stop using that feature.
When the decision affects your work, education, care or access to a service, seek the responsible person and the review route. The machine may have assisted, but an organization remains responsible for how it uses the result. Keep a record of the specific error and its effect rather than guessing at the model’s internal motives.
Look for the missing person in the process
A fairness report can look reassuring while leaving out the people who could not complete the process. If a user abandons an inaccessible form before the model sees their application, a model-only audit misses that exclusion. If someone never appeals because the route is hidden or intimidating, the absence of complaints says little about the outcome.
Map the journey from first contact to final decision. Who hears about the service? Who can submit information in a usable format? Who can see and correct what the system inferred? Who gets a meaningful explanation? Who has the power to change an error? A fairer model inside an unfair process is not enough.
This is where cross-functional work matters. Designers, content writers, data specialists, accessibility testers, subject experts and affected users each notice different problems. Make room for disagreement and document the trade-offs. A person should not have to prove that their way of thinking is typical before the system treats them with care.
Questions people ask
Does AI always discriminate against neurodivergent people?
No. A particular outcome depends on the tool and setting. AI can be useful, but teams must test for exclusion, accessibility and harmful assumptions in the actual workflow.
Can removing a diagnosis from the data eliminate bias?
Not by itself. Other features and historical labels may still carry patterns. Privacy matters, and so do context, evaluation, user participation and human accountability.
Is a human reviewer enough to make an AI decision fair?
Only if that person can understand, question and override the recommendation. A nominal reviewer who simply accepts a score does not solve the problem.
How can I use AI without losing my own voice?
Give the tool a specific task, keep the original, inspect each change and reject suggestions that erase meaning or style. Verify important claims with their source.
Keep exploring
For the broader mechanics of bias, see How Bias Gets Built Into AI Systems. For the interface side, watch Designing Digital Spaces for Neurodivergent Individuals. For a more personal question, read How AI Can Affect Mental Health and Emotional Well-Being.
The question is not whether a machine can produce an impressive answer. It is whether the people affected can understand it, use it, correct it and remain themselves while doing so.
Go deeper
Read Feeling Seen by a Machine online, or keep the complete edition on Kindle or in paperback.
Read the book