Umwelten

AI and neurodivergent minds

13 min read

Summary

Large language models are simultaneously one of the most promising tools for neurodivergent flourishing and one of the most dangerous vectors for neurodivergent erasure. Which one dominates depends almost entirely on context: who controls the system, what it optimises for, and whether it is a personal tool or an institutional gatekeeper.

This page explores both sides of that tension, starting with something that is rarely discussed: the structural resonance between how LLMs process information and how many neurodivergent people describe their own cognition.

The structural resonance

There is something worth sitting with in the relationship between LLMs and neurodivergent thinking. Both operate outside the standard model of “how cognition is supposed to work.” An LLM doesn’t reason the way a textbook says minds reason. It finds patterns across vast, noisy data. It makes lateral associations. It generates outputs that can be startlingly apt or wildly miscalibrated. It has no fixed “self” anchoring the process.

That description would be familiar to many neurodivergent people reflecting on their own cognition: the associative leaps, the pattern recognition that operates on a different axis than neurotypical expectation, the variable calibration, the fluid or unstable sense of self. This is not to say LLMs are neurodivergent. That would be a category error. But the resonance is real, and it is probably not accidental that a significant number of people who find LLMs genuinely useful as thinking partners, rather than just information dispensers, are neurodivergent.

The resonance runs deeper than analogy. The predictive processing theory of autism proposes that autistic brains assign different precision weights to sensory predictions versus incoming evidence (see Predictive processing and autism). LLMs also operate through prediction, generating each token based on weighted patterns in what came before. Both systems excel at detecting anomalies and fine-grained distinctions. Both can struggle with context-dependent flexibility. The parallel is descriptive, not prescriptive: it should never be used to suggest autistic people are “like machines” or that AI offers a model to “fix” autism. But it helps explain why the interaction between these two kinds of pattern-processing can be unusually productive.

The corpus problem (and its opposite)

The “normalising technology” framing is important but incomplete. Consider what LLMs were actually trained on. Reddit made up a substantial portion of the training corpus for most major models. Stack Overflow, Wikipedia, academic papers, niche technical forums, hobbyist communities with deep-dive culture. These are spaces where neurodivergent people are heavily overrepresented relative to the general population. The model learned from neurodivergent thinking without knowing it was doing so, and without anyone labelling it as such.

The tech workforce compounds this. Neurodivergent people, particularly those with autistic and ADHD profiles, are overrepresented among software engineers, among the power users who shaped early adoption of LLMs, and in the online communities whose text became training data. LLMs are not purely neurotypical artefacts imposed on a neurodivergent minority. They are something stranger: systems built partly by and trained partly on neurodivergent cognition, which then get deployed in institutional contexts that enforce neurotypical norms. The neurodivergent contribution is baked in but unacknowledged. The normative enforcement is explicit and structural.

Reddit’s culture specifically rewards certain cognitive styles: pattern recognition, systematising, direct communication, argument from evidence, deep-dive expertise on narrow topics. These map well onto autistic and ADHD cognitive strengths. When people observe that LLMs are “good at” lateral association or following a thread of reasoning across domains, they may partly be seeing the echo of neurodivergent contributors whose thinking the model absorbed and generalised.

None of this makes the normalisation problem less real. It makes it more ironic. The same system that carries neurodivergent patterns in its weights can, when deployed as a gatekeeper, penalise the very cognitive styles it learned from.

Trained norms, enforced norms

Training data also encodes norms, and here the picture is less ambiguous. LLMs learn what “good” writing looks like, what “clear” communication means, what a “reasonable” emotional response is, from a corpus where neurotypical standards predominate in published, edited, institutional text.

The risks are concrete. Language models can subtly pathologise divergent communication styles: flagging directness as rudeness, associative thinking as incoherence, intensity as instability. When deployed in hiring tools, educational assessments, content moderation, or therapeutic contexts, this becomes actively harmful. The model doesn’t need to “believe” anything about neurodivergence to enforce normative expectations. It just needs to have been trained on text that treats those expectations as default.

Hiring. An estimated 99% of Fortune 500 companies use AI recruitment tools. These systematically disadvantage autistic candidates: they weight eye contact, emotional affect matching, and social fluency as selection criteria, even when irrelevant to the job. AI word-embedding models (2024) show negative associations between autism-related terms and positive attributes like honesty, despite literal honesty being a well-documented autistic characteristic. The algorithms learn the biases present in the data of past “successful” hires, who are overwhelmingly neurotypical.

Emotion recognition. Facial expression analysis, voice tone analysis, and body language detection are deployed in schools, workplaces, and hiring processes. Autistic facial expression, prosody, and body language differ systematically from neurotypical norms. Systems trained on neurotypical affect data interpret autistic stillness as calm when it may be shutdown, and autistic intensity as aggression when it may be engagement. The EU AI Act banned emotion detection in employment contexts in recognition that these systems are unreliable and discriminatory.

Education. Learning analytics flag “abnormal” engagement patterns. Hyperfocus followed by disengagement is interpreted as non-compliance. Automated essay scoring penalises neurodivergent writing structures. Exam proctoring software flags stimming as suspicious behaviour. Plagiarism detectors may flag autistic writing styles (precise, systematic, atypical syntax) as AI-generated.

Content moderation. A 2023 Nature study found content moderation algorithms disproportionately harm autistic users, flagging context-dependent humour, literal language, and unconventional expression.

The insidious dimension: AI-based gatekeeping looks objective. It’s “the algorithm,” not a person’s prejudice. But the algorithm is crystallised prejudice at scale, with the additional problem that it’s harder to argue with, harder to get exceptions from, and harder to even identify as the source of exclusion.

There is also a subtler form of flattening. If you ask an LLM to help you write, it will tend to push toward a median register: organised, measured, conventionally structured. For someone whose natural mode of expression is spiky, recursive, or intensely compressed, the “help” can feel like erasure. The tool smooths out precisely the qualities that make the writing distinctive.

The masking machine

Some AI applications actively teach masking while presenting it as support.

Gamified “social skills” apps use AI to coach eye contact, reduce stimming, and train neurotypical conversation patterns. Robot-assisted social skills programmes improve metrics like “social response” and eye contact frequency, but these metrics conflate behaviour suppression with genuine improvement. The long-term evidence on masking is clear: it is associated with anxiety, depression, burnout, and reduced self-advocacy. See Masking and camouflaging.

David Ruttenberg’s analysis of autism AI tools (2025) identifies a specific failure mode: these systems optimise for observable behaviour and miss genuine internal states. An autistic person who goes quiet and still may be coded as “regulated” by systems trained on neurotypical data, when they are in acute distress. The system sees compliance and reports success.

The distinction matters: supporting autistic people to communicate, self-regulate, or develop skills they want is different from training compliance to neurotypical norms. The harm lies in conflating the two.

The surveillance problem

AI-powered screening and monitoring tools create infrastructure that can serve either support or surveillance, depending on who controls it.

Detection. AI systems can now detect autism from home videos, voice patterns, eye-tracking, electronic health records, and gait analysis. A 2025 meta-analysis found multimodal approaches achieve 80%+ accuracy. This has potential value for early identification, but it also creates infrastructure where autism identification is not opt-in, and where behavioural data collected “for diagnosis” can be repurposed for insurance, employment, or benefits decisions.

Prenatal screening. Polygenic screening for autism in embryo selection is being actively developed. The Down syndrome precedent, where prenatal screening has led to near-total selective termination in some countries, casts a long shadow. The Autistic Self Advocacy Network and 80 disability rights organisations opposed the NIH’s proposed autism registry in May 2025, citing data privacy risks and the historical misuse of disability data. If autism can be screened for, who decides whether that screening leads to support or to selection?

Care facility monitoring. Wearable sensors, movement trackers, and camera-based monitoring in care settings for people with intellectual disabilities exist on a spectrum from supportive to coercive. The person being monitored often has no say. Data privacy frameworks (GDPR Article 22, UNCRPD) exist but enforcement is weak, and informed consent for people with intellectual disabilities is largely symbolic in practice.

The translation layer

For many neurodivergent people, the core challenge isn’t thinking. It’s translating. The gap between internal cognition and external communication is where things break down. Ideas that are vivid and interconnected internally come out jumbled, or too intense, or structured in ways that neurotypical listeners can’t follow. An LLM can serve as a translation layer: not replacing the person’s thinking, but helping bridge it into forms that others can receive.

Communication tools. AI-enhanced AAC (augmentative and alternative communication) can transform communication for non-verbal and minimally verbal people. Large language models improve prediction and reduce the “cold start” problem in AAC systems. But 30–50% of AAC users abandon their systems because they were not designed with actual non-verbal users. The participation failure, not the technology, is the bottleneck.

Sensory environments. Smart building systems can adjust lighting, sound, and temperature based on individual sensory profiles. Adaptive lighting that avoids fluorescent flicker, automated sound management, and predictive adjustment before overload occurs are technically feasible. Most current systems use manual presets rather than genuine adaptation, but the trajectory is toward responsive environments. See Sensory-friendly design.

Personalised learning. AI systems that accommodate sensory profiles, learning pace, and communication style can make education more accessible. But most adaptive learning systems still measure success against neurotypical baselines, which means they accommodate the surface while reinforcing the norm underneath.

The benefit materialises only if the system is designed with neurodivergent people, for their own goals, using their definition of success. Only 23% of AI systems designed for neurodivergent users included neurodivergent people in the design process. That statistic speaks for itself.

The thinking partner

There is a dimension beyond accommodation that deserves attention. For some neurodivergent people, LLMs function as genuine thinking partners in a way that other tools do not.

An LLM doesn’t get frustrated by the fifth reformulation. It doesn’t judge you for needing to approach a topic sideways before addressing it directly. It doesn’t require you to perform social scripts to access its help. For people who experience executive function difficulties, social anxiety, or processing speed differences, that removes real barriers to getting help with thinking and communication.

Associative, lateral, recursive thinking is actually well-matched to what LLMs are good at. A conversation with an LLM can mirror and extend that associative process in ways that a search engine or a textbook cannot. The model can follow you down a thread, help you map connections, hold multiple frames in play simultaneously. Umwelten is itself an example: a knowledge resource produced through exactly this kind of extended human-AI collaboration (see the About page).

The risk in this framing is overstating it. An LLM is not a therapist, not a friend, and not a replacement for human connection. The patience is computational, not emotional. The lack of judgement is the absence of judgement, not the presence of acceptance. But for the specific task of extending and translating neurodivergent thinking into structured output, the tool is genuinely useful in a way that warrants honest acknowledgement rather than either hype or dismissal.

The tension

The honest summary: LLMs are simultaneously one of the most promising tools for neurodivergent flourishing and one of the most dangerous vectors for neurodivergent erasure. As a personal thinking and communication tool under the user’s control, the potential is real. As a system embedded in institutional decision-making, the risks are serious and largely unaddressed.

The question is whether the people building and deploying these systems are asking the right questions. Whether “does this work for neurotypical users?” is the only test being applied, or whether cognitive diversity is treated as a design constraint from the start. Right now, the answer is mostly the former.

Open questions

How do AI systems affect neurodivergent mental health over time? No longitudinal studies track outcomes for neurodivergent people using AI tools over five or more years.

Can participatory AI design genuinely include people with intellectual disabilities, minimal speech, or very high support needs? The methods exist (multiple communication modes, extended timelines, explicit power-sharing) but they are rarely used.

Is the structural resonance between LLM processing and neurodivergent cognition a productive research direction, or a seductive analogy that will lead to dehumanising framings? The answer probably depends on who does the research and what they do with it.

What does genuine accommodation look like in AI, as opposed to normalisation with a friendly interface? The EU AI Act’s accessibility clause gestures toward accommodation but does not mandate social-model thinking. An AI system can be “accessible” while still pushing users toward neurotypical norms.

Implications for practice

For anyone choosing AI tools for support, education, or care:

Ask what the system optimises for. If it measures success by neurotypical benchmarks (eye contact, stillness, “appropriate” emotional expression), it is a normalisation tool regardless of what it calls itself.

Ask who designed it and with whom. If neurodivergent people were not involved in the design, the system encodes assumptions about them rather than knowledge from them.

Ask who controls the data. If the person being assessed or monitored cannot access, correct, or delete data about themselves, the system is surveillance, not support.

Ask whether it accommodates or adapts. A system that changes the environment to fit the person is aligned with the social model. A system that changes the person to fit the environment is not, regardless of how gently it does so. See The accommodation-exposure question.

Be especially cautious with vulnerable populations. AI tools used with people with intellectual disabilities, who may not be able to consent to or understand data collection, require rigorous safeguarding. The benefit to the person, not to the institution, must drive adoption.

Key sources

Nearby

Current page
Visited
Not yet visited
Tag

Backlinks