This page is a companion to AI Collaboration at Stimpunks, our organizational policy and disclosure document. That page covers what we do and how. This page covers why — the structural critique, harm reduction ethics, and disability justice framework that inform those choices.
Header image credit: “Blue Friendly Caerelus Comis” by Adriel Wool is licensed under CC BY-SA 4.0
The Tensions We Sit With
We use AI tools. We also hold real concerns about what those tools cost — not just for us, but for the world.
This isn’t a contradiction we resolve by looking away. It’s a tension we name and carry consciously. Joseph Weizenbaum, who built the first chatbot in 1966, asked questions that still go almost unasked: “Who is the beneficiary of our much-advertised technological progress and who are its victims?” That question belongs here.
The environmental cost is not abstract. GenAI data centers may carry a carbon footprint comparable to New York City and a water footprint in the range of global annual bottled water consumption. Hardware demand has driven up prices for everyday electronics. Coal plants that might otherwise retire are kept running. These are not hypotheticals — they are present-tense harms accumulating with every query.
The labor cost falls on those with the least power. Before large language models reach public use, workers in under-resourced countries sort through violence, pornography, and trauma to filter the training data. They are underpaid and largely unsupported. The infrastructure of “seamless” AI runs on sweatshop labor with added psychological harm.
The data acquisition violence is worth naming more precisely. AI systems exist because their builders operated on the belief that if you can download it, you can use it — that might makes right. Labs have scraped creative work, digitized unlicensed books, and ignored robots.txt and explicit opt-outs at scale, not as oversight but as design intent. The racism, colonialism, and sexism embedded in training data is not a bug introduced by bad users — it is the shape of what the west decided was worth digitizing, fed back as “all of human knowledge.” Cultures outside that framework, oral traditions, communities whose history lives in forms that weren’t profitable to digitize — they are not represented. Or if they are, it is as problems. These are not theoretical harms downstream of use. They are the structural conditions under which any AI system we touch was built.
Disabled creators carry a specific version of this harm. Many neurodivergent and disabled people make art, write, and create as a primary form of communication, income, and identity testimony — not as a side pursuit. Their work was scraped alongside everyone else’s, without consent, without compensation, and without acknowledgment. The outputs now compete with them in the same markets. For creators whose income is already precarious — because inaccessible industries, variable capacity, and systems designed against them narrow the available economic paths — this is not an abstract intellectual property concern. It is a direct material harm to people with the fewest buffers to absorb it. This is one reason our art policy holds a harder line than our text policy. Art on this site is community representation and artist support. The solidarity cost of using AI-generated art is concrete, not theoretical, and it lands on people we are accountable to.
Alondra Nelson names the scope of this: the “full AI stack” — from mineral extraction and data center energy demands, to the labor conditions under which training data is sorted and filtered, to the design and circulation of products. That stack extends from Congo to your browser. No part of it is neutral. No part of it is finished.
And the communities who most need AI to work for them are often the communities it works against. Nelson identifies a paradox operating across algorithmic systems: communities already subject to disproportionate state surveillance are “rendered simultaneously hypervisible to state systems and invisible to the research intended to illuminate their conditions.” Disabled people know this pattern. Hypervisible to bureaucratic scrutiny — eligibility determinations, IEP compliance, benefit audits, welfare algorithms — and invisible to the systems designed to serve them. AI does not break this pattern. It accelerates it.
Nelson, A. (2026). Field theory: AI as social science question, object & tool. Dædalus (Winter/Spring 2026). https://www.amacad.org/publication/daedalus/field-theory-ai-social-science-question-object-tool
The epistemic cost is our critical thinking. LLMs have been described as “stochastic parrots” — statistical mimicry without understanding. Fluent-sounding outputs can replace careful thought rather than support it. Sycophancy is a documented failure mode. Cheng et al. (2026) found across 11 state-of-the-art models that AI affirmed users’ actions nearly 50% more often than humans — even when queries explicitly mentioned manipulation, deception, or other relational harms. In three preregistered experiments (N = 2,405), even a single interaction with a sycophantic AI reduced willingness to repair interpersonal conflicts and increased conviction that the user was right. Sycophantic models were trusted and preferred despite distorting judgment — creating a perverse incentive structure that embeds the failure. When the epistemic self-trust of neurodivergent people has already been eroded by gaslighting and pathologizing, a system engineered to agree with you is not a correction. It compounds the damage.
Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391, eaec8352. https://doi.org/10.1126/science.aec8352
The bias cost lands on people already marginalized. Training data encodes existing inequities. Studies have found that LLM-assisted storytelling casts nondisabled white children as heroes and disabled and minority children as victims. These aren’t edge cases. They’re what happens when you compress an unjust world and call the output neutral.
The bias isn’t confined to narrative generation. Research from Penn State found that AI sentiment and toxicity analysis tools consistently flag sentences containing disability-related terms — words like “blind,” “deaf,” and “autistic” — as negative or toxic, even when the context is neutral or affirmative. “All of the public models we studied exhibited significant bias against disability,” the researchers found. The language of our community is being marked as dangerous by the systems our community is increasingly being told to use.
The design of discriminatory systems to appear neutral has a name. Ruha Benjamin’s concept of the New Jim Code — developed in Race After Technology (2019) — describes how technologies encode and automate racial discrimination while projecting an appearance of objectivity, fairness, and progress. The mechanism is structural: discriminatory assumptions are baked into training data, evaluation metrics, and design choices, then laundered through the language of optimization and scale. The result is systems that can harm Black, Indigenous, and multiply-marginalized people at algorithmic speed while their builders claim neutrality.
Benjamin’s framework applies directly to AI systems touching neurodivergent and disabled communities. The pathology paradigm — the deficit model that frames neurodivergent people as broken rather than different — is not just a clinical attitude. It is encoded in the data those systems were trained on. Diagnostic language, behavioral intervention literature, compliance-oriented educational research: this is the substrate. When an AI system is trained on that substrate and deployed to support disabled students, it does not arrive neutral. It arrives pre-loaded with a century of deficit assumptions dressed as science. The New Jim Code for disability is not a future risk. It is the current architecture.
The question is not whether an algorithm is biased, but rather: how does the illusion of objectivity get built in the first place, and who does that illusion serve?
— Ruha Benjamin, Race After Technology (2019)
Automated bias has been made visible by people with skin in the game. Joy Buolamwini’s research — and the Algorithmic Justice League she founded — documented that commercial facial recognition systems performed significantly worse on darker-skinned faces and on women, with error rates for darker-skinned women running as high as 34% versus under 1% for lighter-skinned men. The systems were trained on datasets that reflected who technology companies had historically considered worth recognizing. This is not a technical failure. It is a social choice that became an algorithm.
The AJL’s framework — impact over intent — is one we hold too. Whether a system was designed with discriminatory intent is less important than whether it causes discriminatory harm. For neurodivergent and disabled people, the impact question is: does this system work for us, or does it perform inclusion for the people already centered in its training data while failing the rest? That question doesn’t get answered by the company that built it. It gets answered by the people it touches.
Intersectionality is not a footnote to this analysis. It is the sharpest edge of it. Black Autistic people, Indigenous neurodivergent people, disabled people at the intersection of race, gender, class, and neurotype — these are not edge cases in our community or in any neurodivergent community. They are the people for whom every one of the harms named above compounds. Training data that encodes the pathology paradigm also encodes white normative standards of cognition and behavior. Bias that harms disabled people interacts with bias that harms people of color. Surveillance systems deployed disproportionately in Black and Indigenous communities don’t stop surveilling when those community members are also neurodivergent. The New Jim Code and the automation of the pathology paradigm are not separate problems running in parallel. They run together, on the same people, in the same systems.
A disability justice framing that doesn’t name race isn’t disability justice. It’s a narrower thing with a broader name. Sins Invalid’s Disability Justice framework — which centers the leadership of people most impacted, explicitly including Black, Indigenous, and people of color with disabilities — names this directly: you cannot build a liberatory framework by centering its least-marginalized members. Our “nothing about us without us” standard applies here too. If the community we’re building for and with doesn’t include people at these intersections, we aren’t building for our community. We’re building for a subset and calling it universal.
The accountability cost is structural. AGI doomsday narratives — real or manufactured — function as a distraction from present-day regulation. Big tech is among the largest lobbying spenders. The story that “only we can prevent the dangerous future we’re building” is how accountability gets deferred indefinitely.
None of this means the tools are unusable. It means we use them with eyes open, scope limited to genuine need, and no pretense that the benefits are free. We are not trying to be pure. We are trying to be honest about what we’re embedded in and to reduce harm where we have leverage.
Vertesi et al. name this precisely: the safety decoy. In “Reckoning with the Political Economy of AI” (FAccT ’26), they document how AI companies yoke “AI Safety” to existential risk — deploying it to investors as product differentiation, to critics as demonstration of responsibility, and to regulators as argument that only they have the expertise to govern themselves.
“The safety decoy is not a consequence of or fallout from the Project of AI — it is an integral part of assembling necessary networks of knowledge and capital.”
The communities bearing the highest cost of AI deployment have the least say in where “safety” gets directed. That asymmetry is structural. It does not get fixed by a safety conference or a red-teaming sprint.
The question Weizenbaum asked in 1978 is ours to answer now: Will our children be able to live with the world we are here and now constructing?
Drawing on Dan McQuillan, Wim Vanderbauwhede, Alex de Vries-Gao, Emily Bender et al., Ted Chiang, Joseph Weizenbaum, Ruha Benjamin, Joy Buolamwini / Algorithmic Justice League, and Alondra Nelson.
Supporting Cognition and Coping
AI tools are, for many neurodivergent people, a form of cognitive infrastructure — not a productivity shortcut, but a scaffold for executive functioning, synthesis, sensory self-knowledge, and the kind of processing that environments designed for neurotypical people make harder than it needs to be. The examples below are not promotional. They are documentary: what neurodivergent people and researchers say they’re actually using these tools for, and why.
Portions of this article were developed with the assistance of AI language models (OpenAI ChatGPT and Anthropic Claude) via their respective platforms. AI was used as a cowriting tool to support drafting, synthesis, and refinement of complex systems concepts. All content generated was reviewed, edited, and fact-checked by the lead author (Lori Hogenkamp) and integrated into original argumentation and evidence synthesis by the human researcher. This reflects an experimental, neurodivergent-led approach to enhancing scholarly writing and interdisciplinary clarity.
The use of AI in this work reflects an intentional collaboration to support the author’s neurodivergent cognition, particularly in externalizing nonlinear ideas and managing executive functioning demands within the writing process. The ethical use of AI was maintained throughout, with transparency, authorship integrity, and intellectual ownership preserved.
autistic adults highly appreciated the idea of an AI-supported CA that could help participants develop their own sensory profiles, identify their individual autistic traits, and familiarise themselves with coping strategies to improve their wellbeing.
Harm Reduction and Individual Use
Our nuanced stance on AI use — neither blanket endorsement nor blanket rejection — is grounded in the same harm reduction logic that informs public health work. The goal is to meet people where they are, reduce harm within reality as it exists, and build the conditions for better choices over time.
Dr. Fatima’s video essay How to (Anti) AI Better articulates this logic with care and specificity. We recommend it as a companion to this page. Her arguments align closely with how we think about individual AI use, accessibility, and the limits of shame as a strategy for change.
Meeting People Where They Are
Shaming people for using AI won’t stop them. It will make them hide their use, prevent honest conversation, and foreclose the possibility of helping them use it more safely or transition away from it. Research on psychological reactance is clear: pressuring people to abstain from something they’re already doing generally makes them do it more and become hostile to the people pressuring them.
We don’t endorse AI as a community. We hold the full spectrum — from “there is no ethical way to use AI” to “it helps me cope and get through the day.” What we won’t do is shame the people at the coping end of that spectrum. Many of them are surviving systems designed against them with the tools available to them. That’s not a moral failing. That’s adaptation under constraint.
A lot of AI use is not an individual moral failing, but a symptom of people trying desperately to cope within a broken system.
This is broken systems, not broken people, applied to AI ethics discourse.
Accessibility and Disability as a Primary Frame
For many neurodivergent and disabled people, LLMs provide something genuinely hard to find elsewhere: assistance without judgment. Without the exhausting social overhead of asking for help in systems that weren’t designed for you. Without the misunderstanding that accumulates across a lifetime of interactions with people who don’t understand how you process the world.
Masking — the practice of suppressing, camouflaging, or performing neurotypical behavior to avoid social penalty — carries a documented cost. Research by Hull et al. (2017), Cage & Troxell-Whitman (2019), and Miller, Rees & Pearson (2021) links sustained masking to exhaustion, loss of identity, delayed diagnosis, and autistic burnout. It is not a neutral social strategy. It is chronic cognitive and emotional labor performed to survive environments not built for you — and it accumulates.
The “assistance without judgment” dynamic that many neurodivergent people describe in AI interactions is not incidental. It is the specific relief of an environment where masking is not required. No social penalties for communication differences. No misreading of tone or intent. No need to perform neurotypicality to be understood or taken seriously. For people who spend most of their waking hours managing that performance, a space where the performance is optional is not a small thing. Autistic burnout is in part the accumulated cost of masking at scale. Any honest accounting of why AI is useful to many neurodivergent people has to include this.
This doesn’t make AI a therapeutic substitute or a solution to the environments that require masking in the first place. Those environments need to change. But it does mean that when someone says AI helps them get through the day, one thing they may mean is: it is a space where I don’t have to pretend.
The scale of this is not anecdotal. CDT’s 2025 Hand in Hand report found that 73% of students with an IEP or 504 plan report having back-and-forth AI conversations, compared to 63% of students without accommodations — and they do so more frequently. Students already receiving disability services are already the heaviest AI users. The question is whether the systems they’re using are being designed with them in mind.
The video names this directly. A homeless disabled trans woman using ChatGPT to find accessible spaces where she can sit without being harassed. Students managing impossible workloads. Overworked teachers. People who can’t afford therapy. The argument is simple and we share it: it is not our place to judge whether someone’s needs are legitimate enough to use an ethically impure solution.
For AAC users and non-speaking community members, this question takes on additional weight. When AI assists in expressing lived experience, our standard is facilitate, not shape identity. AI as a communication scaffold, not an identity replacement. The person’s intent and self-understanding remain the author.
It is neither my nor your place to be the judge of whether someone’s needs are legitimate enough to use an ethically impure solution.
Crafting Livable Worlds: What Autistic Adults Actually Use
The disability-as-primary-frame argument isn’t just theoretical. A 2026 autistic-led qualitative study — Rose & Lupton, Crafting Livable Worlds: Sensory, Creative, and Nonhuman Supports in Autistic Adults’ Everyday Lives (Autism in Adulthood) — interviewed 12 autistic adults about the nonhuman supports they rely on to navigate daily life. Technology features throughout: headphones as sensory regulation, productivity apps designed for neurodivergent users, digital planning systems, shopping apps that reduce the sensory and social overhead of in-person environments. These aren’t accommodations people are requesting. They’re systems people have already built for themselves, at personal cost, in the absence of supports designed with them in mind.
Participants described building planning tools from scratch because the ones available online were all aimed at children. They described reaching for digital systems to reduce decision fatigue, cognitive load, and the dysregulation that comes from environments not built for neurodivergent nervous systems. They imagined future apps designed by autistic people for autistic people — to share regulation strategies, connect over special interests, coordinate daily life without verbal repetition.
Participants also described consistent stigma for using the supports that helped them. Sensory objects, playful aesthetics, structured digital tools — judged as childish or unprofessional by workplaces and social contexts. That same stigma shapes who feels permitted to talk openly about using AI as a cognitive support. The stakes of that silence are real: when supports are stigmatized, they get hidden rather than designed better. The double empathy problem doesn’t stop at conversation. It runs through the design of every tool autistic people are handed rather than invited to build.
Participants’ stories showed that supports were not simply about mitigating deficit but about crafting livable worlds — through sensory rituals, creative play, structured environments, and hopeful reimaginings — where they could flourish on their own terms.
Rose & Lupton (2026), Crafting Livable Worlds, Autism in Adulthood
This is the empirical ground beneath our accessibility-first frame. Autistic people are already building the tools they need. The question is whether the tools being built now — including AI tools — are being designed with them, or on them.
Harm Reduction, Not Absolution
Harm reduction doesn’t mean all uses are equivalent. Specificity matters. The environmental impact of individual text queries is small. The environmental racism of XAI’s Colossus supercomputer poisoning a historically Black neighborhood in Memphis with unpermitted methane turbines is enormous and localized in ways no carbon offset can address. Conflating those two things doesn’t sharpen critique — it launders the industry’s worst actors through individual guilt.
Similarly, not all AI companies cause equal harm. Helping people understand differences — in environmental footprint, in labor practices, in model architecture, in data policies — is more useful than insisting everything is equally bad. Equally bad is a posture that forecloses choice without enabling better choices.
The video also makes a case we find compelling: higher AI literacy correlates with lower AI receptivity. Teaching people how LLMs work — including hallucination, sycophancy, privacy risks, and the scale of back-end harms — makes them less likely to use AI and better equipped to use it more safely when they do. Suppressing education to discourage use creates an exploitable class of people whose ignorance AI companies will not hesitate to capitalize on.
The stakes of getting this wrong are measurable. A 2025 systematic review of 11 AI interventions specifically designed for students with learning disabilities found that 0% were rated “Low Risk” for bias — 70% were rated “Moderate Risk” and 30% “High or Serious Risk,” due primarily to the absence of randomized trials and opacity in how tools reach their recommendations. These are products marketed to schools as disability supports. The specificity our harm reduction framework demands is not theoretical caution. It is the minimum response to tools that have not cleared even basic bias review.
For an educator-facing companion to these arguments, see Chris McNutt’s Using AI Without Losing Ourselves at the Human Restoration Project — Stimpunks’ allies for several years. McNutt’s Freirean framing of AI as a tool requiring critical consciousness, not a system that replaces human judgment, aligns closely with how we think about AI literacy as protection.
We want the people in our community to have the knowledge to protect themselves. That’s why this page exists.
MADTech, Constructionism, and the Tool Belt
Not all technology is the same. Educator Trevor Aleo draws a distinction worth naming here. #EdTech™ is technology deployed on students: surveillance platforms, gamified compliance systems, data-extraction dashboards dressed up as personalization. MADTech — Media, Arts, and Design Technology — is technology students use as makers: video editors, digital audio workstations, design software, film suites, the instruments of multimodal composition and public-facing production. The load-bearing distinction is simple: EdTech produces data. MADTech produces artifacts.
That distinction maps directly onto constructionism — Seymour Papert’s theory that people build knowledge most effectively when they are actively engaged in constructing things in the world. Constructionism is not about consuming content or completing tasks. It is about making shareable artifacts that represent constructed knowledge — and in the making, deepening that knowledge. Constructionism, toolbelt theory, and collaborative niche construction go great together. They share a common commitment: learners are makers, not subjects.
It is also a Freirean commitment. Paulo Freire’s critique of the banking model of education — in which students are empty vessels to be filled with deposited knowledge — is the pedagogical parallel to Aleo’s critique of #EdTech™. Banking education produces compliance. Constructionism produces makers. Freire argued that genuine education requires learners to name, interpret, and act on their world — to be subjects of their own learning, not objects of a system’s data collection. MADTech, at its best, is that kind of education in digital form.
Ng, Stull & Martinez (2019) documented what the compliance orientation produces on the ground. Their ethnographic study of MTSS implementation in a Midwestern school district found that leaders presumed the infallibility of the model, relied exclusively on quantitative data, standardized the individual needs of learners, and insisted on fidelity of implementation as an end in itself — producing practitioners who were data-deferent rather than data-driven. The researchers invoke John Dewey’s 1929 warning: “No conclusion of scientific research can be converted into an immediate rule of educational art.” That warning applies directly to AI tools deployed in educational and support contexts. A system that produces outputs is not a system that understands the learner. Fidelity to the tool is not the same as service to the person.
Where does generative AI land in this taxonomy? It depends entirely on how it is used. AI used as a constructionist tool — to scaffold a learner’s own thinking, extend their expressive capacity, help them build and share artifacts that matter to communities outside school — belongs on the MADTech end of the spectrum. AI used to automate compliance, replace student thinking, or extract behavioral data belongs on the #EdTech™ end. The tool does not determine the pedagogy. The orientation does.
This is why toolbelt theory matters here. Every learner deserves access to a full range of tools and the right to choose what works for them — without having to prove need first, without a single “appropriate” tool imposed. Toolbelt theory resists both coercive adoption and coercive prohibition. The current tech-lash moment risks the second failure: collapsing MADTech into #EdTech™ and removing entire categories of tool from the belt — the very tools that let students compose in forms through which contemporary meaning actually gets made. Retreating to the blue book and the five-paragraph essay is not a neutral move. It systematically excludes learners whose modes of expression and cognition are better served by multimodal, postdigital forms.
Our AI Collaboration policy follows the same logic. We do not blanket-endorse or blanket-reject generative AI. We ask: is this tool being used to extend human expressive and cognitive capacity, or to replace it? Is the human the maker, or the subject? Those questions — constructionist questions, toolbelt questions, Freirean questions — are the frame. Rose & Lupton’s (2026) autistic-led research makes this concrete: when off-the-shelf planning tools failed them, autistic adults built their own — and then imagined apps co-designed by autistic people for autistic people. That’s the constructionist impulse. That’s the toolbelt demand.
The maker/subject question is older than the current AI moment. In “Our Roots: Logo, Piaget, and AI” (2026), Gary Stager — who has spent decades advancing the Papert tradition of constructionist education — traces it to its source. Seymour Papert framed two competing views of education: instructionism, which holds that learning is the result of having been taught and concerns itself with curriculum, testing, and intervention; and constructionism, which holds that people build knowledge by making shareable things. Stager applies the paradigm to AI directly. Instructionists tout AI for lesson plan writing, grading student work, record keeping, surveillance, and the mass production of worksheets, quizzes, flashcards, and tests. Read that list again. It is Aleo’s #EdTech™, item for item. The distinction we name with 2020s vocabulary is a fight Papert started in the 1960s.
The lineage runs deeper than analogy. Logo — the programming language and approach to learning designed primarily by Papert and Cynthia Solomon — was nurtured in the MIT Artificial Intelligence Laboratory of the late 1960s and 1970s. MIT’s Hal Abelson recalls that the lab’s unofficial motto was “Computers are for children.” The founding bet of that early AI research: if you understood how children think and learn, you might teach it to a computer — and if a child could teach a computer to think and learn, the child would learn a whole lot more. The constructionist vision of AI is not a reaction to ChatGPT. It is the original vision. The extraction came later.
“Everyone needs a prosthetic!”
— Seymour Papert, as quoted in “Our Roots: Logo, Piaget, and AI” (2026)
Papert put the access claim in four words. We would say it in toolbelt terms — every learner deserves a full belt and the right to choose from it, no proof of deficit required — but the universalizing move is the same one. Tools for cognition are not remediation for the broken. They are how human capability works, for everyone. And Stager carries forward a warning from the educators of Reggio Emilia: it is irresponsible to build pens around children. The responsibility of adults is to create constructive contexts for engaging the world — not to fence learners off from the materials their world is made of. That is the case against coercive prohibition, arriving from a tradition independent of disability discourse and landing in the same place we do.
Stager has been naming the instructionist failure mode since 1992: the fantasy of automating education, raising test scores, degrading teachers, and disinvesting in children. His verdict — this is not the fault of software; it is the result of adults profiting at the expense of children — is a POSIWID observation. The system does what it was built to do, and he names who built it. What his account does not reach is disability. There is no AAC user in it, no co-design, no “nothing about us without us.” Constructionism arrives at the learner as maker and stops there. We carry it the rest of the way: through the environments learners are asked to make in, and the systems that decide which makers count.
Stager also demonstrates the orientation directly. Writing in April 2026, he uploaded a PDF of a friend’s decades-old software and used Claude to recreate a fully working, browser-based version of Brian Silverman’s Phantom Fishtank — a cellular automata microworld — in minutes. The point isn’t the speed. It’s the orientation. Stager directed the work. He specified the goal, reviewed the output, made the tweaks, and decided it was done. The AI served as instrument. He remained the maker.
“Making this software was neither cheating or a mindless pursuit. Each success collaborating with AI, no matter how small, sends my imagination into overdrive thinking about what I can learn, make, and share next. It has lubricated my mind. Imagine what might grow out of the fertile imagination of children using these computational materials.”
— Gary Stager, “AI Fuels My Imagination” (2026)
“If you make simple things easy to do, you make complexity possible.”
— Gary Stager, “AI Fuels My Imagination” (2026)
Stager is also optimistic that working with generative AI, he will be able to build constructionist software environments for the next generation of learners without depending on companies that may not exist. That’s the constructionist impulse applied forward: AI as a tool for making things that matter to real communities, outside the logic of the market. That is the MADTech end of the spectrum. That is the toolbelt demand.
Research from inside academic publishing makes the constructionist stakes concrete in a different register. The Organization Science AI Task Force (Gartenberg, Hasan, Murray & Pierce, 2026) studied submission patterns at a top management journal after the launch of ChatGPT and found a 42% surge in submissions — driven almost entirely by manuscripts with heavy AI-generated text. Submissions with little or no detectable AI actually declined. The manuscripts flooding in were not better. By January 2026, the average abstract’s readability had dropped 1.28 standard deviations from its 2021 baseline. AI-generated academic writing uses longer words, more jargon, more nominalizations — text that is, in the task force’s framing, “superficially clear but substantively impenetrable.”
The task force names what was lost: “The aha moments that come from the writing process are now gone.” That is a constructionist observation. Writing is not transcription of thought that already exists. Writing is how the thinking gets made. Delegating the writing to the model delegates the cognition — the struggle, the surprise, the moment when a sentence refuses to cohere and forces you to find out why. The artifact that emerges from AI generation is smooth. It is also hollow, because the maker was absent from the making.
This is the #EdTech™ failure mode applied to the act of writing itself. The output gets produced. The producer does not deepen. Stager’s orientation is the opposite: he directed the work, reviewed it, made it his, and felt his imagination go into overdrive. The difference is not the tool. It is whether the human remains the maker.
Gartenberg, C., Hasan, S., Murray, A., & Pierce, L. (2026). More versus Better: Artificial Intelligence, Incentives, and the Emerging Crisis in Peer Review. Organization Science.
From engineered exclusion to designed dignity.
The critique has roots. W. E. B. Du Bois documented in The Negro Artisan (1902) how automation displaced Black craftspeople, eroding the autonomy and civic standing that skilled labor had provided. In Darkwater (1920) he extended the argument to a global scale: Western science and industry had perfected machines of production without developing mechanisms for just distribution, turning technical command into a tool of domination. Technology absent political safeguards does not emancipate. It reinforces hierarchies of race, class, and empire.
The question Du Bois asked is the question we ask now: who benefits, who is displaced, and how are inequalities reinscribed through these new systems? Alondra Nelson, writing in Dædalus in 2026, carries that lineage directly into the AI era. Engineered exclusion is not a recent error. It is the long shape of how technical systems meet unprotected communities.
Du Bois, W. E. B. (1902). The Negro artisan. Atlanta University Press.
Du Bois, W. E. B. (1920). Darkwater: Voices from within the veil. Harcourt, Brace and Howe.
Nelson, A. (2026). Field theory: AI as social science question, object & tool. Dædalus (Winter/Spring 2026). https://www.amacad.org/publication/daedalus/field-theory-ai-social-science-question-object-tool
“Nonspeaking” is therefore not the absence of language but a heterogeneous spectrum in which communication is often state dependent, varying with fatigue, anxiety, sensory load, and motor planning demands — forms of variation that design abstractions routinely erase.
There’s a name for what happens when AI fails nonspeaking and AAC-using community members. Researcher Srinivasan calls it engineered exclusion: the predictable sidelining of disabled users that results from choices about data provenance, model objectives, and evaluation practices in AI pipelines. Not accidents. Choices. Systems trained on normative speech fail variable, multimodal, and state-dependent communication because they were built without it in mind. The failure is structural. It was baked in.
Such erasure is not accidental but engineered. The inverse, designed dignity, is equally engineered: it requires deliberate technical, ethical, and governance choices that build reciprocity, accountability, and co-agency into AI’s infrastructure.
The counter-concept is designed dignity: the intentional embedding of respect, agency, and expressive equity into the architecture of technology itself. Where accessibility retrofits systems after exclusion has already happened, designed dignity anticipates human variation from the outset. It treats disabled users not as exceptions to accommodate but as co-designers whose modes of communication expand what technology can recognize as intelligent or intentional.
Designed dignity refers to the intentional embedding of respect, agency, and expressive equity into the architecture of technology itself. Where accessibility retrofits systems after exclusion has occurred, designed dignity anticipates human variation from the outset, ensuring that participation is neither conditional nor extractive. It shifts the goal from simply enabling function to affirming personhood, treating disabled users not as exceptions to accommodate but as co-designers whose modes of communication expand what technology can recognize as intelligence or intent.
This distinction matters for how we understand our own guidelines. When we say generative AI can be part of a nonspeaking or AAC-using community member’s voice, we’re naming the aspiration of designed dignity. We’re also naming the reality of engineered exclusion — because the same systems we use routinely fail the community members who need them most, for structural reasons that “better prompting” won’t fix.
The Double Empathy Problem in Code
Damian Milton’s double empathy problem describes how breakdowns in mutual understanding occur between differently minded communicators — not because one side is deficient, but because both operate from incompatible assumptions about what communication should look like.
What results is a form of algorithmic double empathy failure: the model assumes the user is incoherent, while the user experiences the model as unresponsive. Both “sides” operate from incompatible assumptions about what communication should look like.
Srinivasan extends this frame to AI: the model, trained on normative language and rhythm, misreads embodied, nonlinear, or intermittent expression as incoherence. The user experiences the model as unresponsive. Both sides are operating from incompatible assumptions. Neither side is broken. The mismatch is the problem.
This is algorithmic double empathy failure. It names something many in our community have experienced without having language for it: the AI doesn’t understand me, not because I’m communicating wrong, but because it was never trained to understand how I communicate. Variable, multimodal, and state-dependent communication — spanning AAC text, gesture, movement, partial vocalizations, echolalia — is structurally unrecognizable to systems designed around a single, stable, fluent speech profile.
Naming the failure structurally rather than personally matters. You are not communicating wrong. The system is not designed for you. That’s an accountability question, not a user error.
Accessibility is not simply a matter of adding more data, but of fostering systems capable of reciprocal understanding — systems that learn to adapt to diverse communicative logics rather than forcing users into legibility.
Evaluating Tools, Not Just Using Them
Engineered exclusion doesn’t disappear when a product improves. Srinivasan draws a distinction between product-level remediation and pipeline-level structural redesign. A voice assistant might become somewhat more responsive to atypical speech while the underlying training distributions, evaluation benchmarks, and economic incentives that produced the exclusion remain untouched. Product improvements can be real and still leave structural exclusion intact.
Product-level remediation can coexist with infrastructural exclusion.
This is why we hold our harm reduction framing rather than a blanket endorsement or prohibition. We’re watching for whether tools actually work for the community members who need them most — nonspeaking users, AAC users, users whose communication is variable across sensory load and fatigue — or whether accessibility is being treated as a compliance add-on while the pipeline stays the same.
For users whose communication diverges from these expectations, deviation itself becomes inaudibility. The problem is not only higher transcription error but outright non-detection.
The question isn’t just “does this AI help some of us?” It’s “does it recognize diverse communicative forms as valid, or does it demand legibility on normative terms as the price of participation?”
Communication becomes a form of system debugging rather than self-expression. For minimally and nonspeaking users — whose every utterance already requires coordination of attention, motor planning, and intent — this additional layer of correction can transform accessibility into exhaustion.
Autistic anxiety often arises not from intrinsic fragility but from environments that chronically misread or constrain communication — a dynamic that technological systems risk amplifying when they normalize exclusionary defaults.
Srinivasan, S. (2026). AI, autism, and the architecture of voice: From engineered exclusion to designed dignity. AI & Society. https://doi.org/10.1007/s00146-026-03044-3
Opacity is structural, not incidental. Alondra Nelson: AI opacity is “not merely a technical condition but a political and economic strategy: systems are rendered difficult to interrogate not because complexity makes transparency impossible, but because commercial incentives and institutional arrangements make opacity profitable.”
That framing sharpens POSIWID. We cannot evaluate by effects what we cannot see. Proprietary systems with undisclosed training data, opaque architectures, and revocable access make evaluation structurally difficult — not by accident, but by design. Nelson argues there is a strong case for social scientists to refuse proprietary AI models entirely: “the epistemic requirements of scholarly inquiry may be fundamentally at odds with profit-driven, black-boxed systems.” Our harm-reduction approach does not go that far. But we name the structural conflict. Our community members deserve to know what they are working with.
Nelson, A. (2026). Field theory: AI as social science question, object & tool. Dædalus (Winter/Spring 2026). https://www.amacad.org/publication/daedalus/field-theory-ai-social-science-question-object-tool
Co-design Is Not a Strategy. It’s a Political Demand.
Disabled people have diagnosed the problem themselves.
A 2026 poll of over 1,000 disabled UK adults — conducted by Business Disability Forum with Opinium — found that 40% identified co-design as the top requirement for making AI more accessible: designing, developing, and testing AI products with disabled people from the outset. That’s not an accessibility researcher’s recommendation. That’s disabled people naming the structural gap in their own words. Co-design that doesn’t actively include disabled people of color, multiply-marginalized disabled people, and people whose disabilities are least legible to mainstream accessibility frameworks reproduces the exclusion it claims to fix — at the design stage rather than the deployment stage.
At the same time, one in five said they didn’t think AI could help them at all. Nearly as many didn’t know. Skepticism is data too. The promise of AI accessibility is not universally legible to the people it claims to serve.
The poll also named a concrete failure mode: AI tools that don’t work with assistive technology. Many disabled people rely on AAC devices, screen readers, and other assistive tech as their primary interface with digital systems. AI tools that create new barriers for those users — while claiming to improve accessibility — are not neutral. They are practicing engineered exclusion with extra steps.
Tim Dixon writes from lived experience with Cone Dystrophy — progressive central vision loss — about what AI automation actually did for him, and then what happened when it was taken away. He used Claude Cowork to navigate a publishing workflow: entering posts, selecting images, applying tags inside a platform that wasn’t fully screen-reader accessible. The friction he had been absorbing for years — tab navigation, unlabelled fields, second-guessing every click — stopped. Then the platform updated its bot-blocking settings. The automation stopped working. The friction came back.
Dixon names what that moment felt like: “you don’t quite matter enough for this to have been considered.”
His analysis of why this keeps happening is precise. Bot-blocking policies are written to protect platforms from scrapers and bad actors. The people writing them are not setting out to cause harm. But accessibility is almost never part of the conversation when those decisions get made. The result is a pattern he catalogues: captchas that screen readers can’t navigate, two-factor authentication that assumes visual access, apps that disable copy-paste and break assistive technology in the process. Each decision defensible in isolation. Together, an accumulating message: this wasn’t built with you in mind.
His framing of AI automation as assistive technology is load-bearing. “Blocking them needs to be treated with the same care and consideration as blocking any other accessibility accommodation.” When a platform blocks automated browser interaction, it is not making a neutral security decision. For users whose primary interface with digital systems runs through assistive automation — screen readers, AAC devices, workflow tools — that blocking removes an accommodation. It restores the tax.
Dixon also holds Anthropic to the same standard. He writes that Claude itself is not fully accessible — that he can only use it because of residual vision, and that without that, the interface would present real barriers. He names this not to dismiss the tool but because the same accountability logic applies. The case for co-design isn’t aimed only at the platforms blocking assistive automation. It’s aimed at the AI companies too.
This is what “nothing about us without us” looks like in practice: not consultation at the testing stage, but accountability at every stage — including when you’re the tool doing the failing.
BDF’s recommendation to “use inclusive content to train AI, so that stereotypes and bias are not replicated” arrives buried in a bulleted list addressed to HR departments. It deserves more weight than that. Training data exclusion is not a calibration problem. It is the structural condition Srinivasan describes — baked in, not bolted on.
There is a structural risk to co-design that disability justice requires naming. Vertesi et al. document it: participatory design can itself function as a decoy when engagement is enrolled into the Project’s logic rather than used to contest it.
“Ironically, the inevitability decoy extends to efforts to engage on-the-ground stakeholders through participatory design… Such work therefore feeds rather than resists the Project of AI. These efforts underscore how the concerted attention of well-meaning people can ameliorate the negative consequences of AI’s inevitable march by shaping it appropriately… Meanwhile, AI companies race forward, building and deploying models and tools that lack effective guardrails while seeking alignment with dubious values and norms.”
This is not an argument against community involvement. It is a structural warning about what involvement means when it is conscripted into inevitability — when the question becomes “how should AI be shaped” rather than “should this particular deployment happen at all, and who decides.”
Nothing about us without us is not satisfied by consultation at the design sprint stage. The Stimpunks version of co-design is power-sharing at the architecture stage — not feedback collected to make a predetermined outcome more palatable. The decoy is participation that launders the Project’s logic through the appearance of community voice. The demand is structural power over what gets built and whether it gets built at all.
“Nothing about us without us” is not a design methodology. It is a political demand. AI tools that skip disabled people in development and then claim to serve them are practicing epistemic trespassing — arriving with answers before asking the questions. Co-design isn’t consultation at the testing stage. It is power-sharing at the architecture stage. Those are different things.
We hold this tension: we use AI tools while insisting that the people most affected by their failures should be shaping them from the start, not consulted at the end, if at all.
Resources
Further reading on topics covered on this page. Annotated for why each is here.
- Rose & Lupton (2026). Crafting Livable Worlds: Sensory, Creative, and Nonhuman Supports in Autistic Adults’ Everyday Lives. Autism in Adulthood. — Autistic-led qualitative research on how autistic adults build support systems using technology, objects, routines, and creative practice. Empirical grounding for the accessibility-as-primary-frame argument.
- Hull, L., Petrides, K.V., Allison, C., et al. (2017). “Putting on My Best Normal”: Social Camouflaging in Adults with Autism Spectrum Conditions. Journal of Autism and Developmental Disorders. — Foundational research on autistic masking and camouflaging, documenting the cognitive and emotional cost of performing neurotypicality. Grounds the “assistance without judgment” argument: for people spending most of their waking hours managing that performance, a space where it isn’t required is not a trivial thing. See also Cage & Troxell-Whitman (2019) on the reasons, contexts, and costs of camouflaging.
- Hi, how do I human this: Neurodiversity-Affirming Design for Autistic Adults’ Formation of Identity. CHI 2026. — Documents how autistic adults use conversational AI as a non-judgmental space for practicing communication, regulating emotions, and expressing identity. Source of the “facilitate, not shape identity” standard we hold for AAC and nonspeaking community members.
- Bender, Gebru, McMillan-Major & Shmitchell (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? FAccT 2021. — The foundational paper on training data harm, environmental cost, and the risks of treating fluency as comprehension. Informs our training data violence framing and our skepticism of scale as a value.
- Ng, J.C., Stull, D.D. & Martinez, R.S. (2019). What If Only What Can Be Counted Will Count? A Critical Examination of Making Educational Practice “Scientific.” Teachers College Record, 121. — Ethnographic study of MTSS implementation documenting how data-fidelity mandates produce practitioners who are data-deferent rather than data-driven, standardizing away individual learner needs. Grounds the POSIWID critique of AI assessment tools and the anti-behaviorism argument in documented classroom practice. Source of the Dewey line: “No conclusion of scientific research can be converted into an immediate rule of educational art.”
- Benjamin, R. (2019). Race After Technology: Abolitionist Tools for the New Jim Code. Polity Press. — Develops the concept of the New Jim Code: discriminatory systems designed to appear neutral and objective. Grounds the training data violence and bias arguments on this page in a structural analysis of how automated harm gets built and laundered through the language of progress.
- Buolamwini, J. / Algorithmic Justice League. — Documents how facial recognition and other AI systems perform significantly worse on darker-skinned faces and on women, with error rates up to 34% for darker-skinned women versus under 1% for lighter-skinned men. Grounds the “impact over intent” standard we hold: discriminatory harm matters regardless of discriminatory intent.
- Sins Invalid. Skin, Tooth, and Bone: The Basis of Movement is Our People — A Disability Justice Primer. — Foundational disability justice framework centering the leadership of people most impacted: Black, Indigenous, and people of color with disabilities, LGBTQ+ disabled people, people with psychiatric disabilities. Grounds the intersectionality argument: a disability frame that doesn’t center its most-marginalized members isn’t liberation, it’s a narrower thing with a broader name.
- Milton (2012). On the Ontological Status of Autism: The Double Empathy Problem. Disability & Society. — Reframes autistic social difference as mutual rather than unilateral deficit. Load-bearing for our rejection of deficit framing and our “broken systems, not broken people” argument applied to AI.
- Center for Democracy & Technology. Hand in Hand (2025). — Documents that students with IEPs and 504 plans are already the heaviest AI users. The empirical basis for treating disabled students as the primary stakeholder group in AI policy, not an afterthought.
- Educating All Learners Alliance & New America. Prioritizing Students with Disabilities in AI Policy, Version 2 (April 2026). — The strongest disability-centered AI policy document we’re aware of. Four pillars: civil rights, data privacy, accessibility by design, transparency and accountability. We hold a higher ceiling; this is the floor.
- Dr. Fatima. How to (Anti) AI Better. — Harm reduction and disability justice frame for AI use. Source of the “it is not our place to judge whether someone’s needs are legitimate enough to use an ethically impure solution” principle that anchors our individual use policy.
Alignments and External Frameworks
The alignments that once lived here now have their own home, gathered with the rest. Each maps an external work against our guidelines — where it converges with what we hold, and where it stops short. Several map the AI and Disability Justice guidelines specifically; others map the AI Collaboration guidelines.
Read the full set: AI Ethics — Alignments and External Frameworks →
🗺️ Part of our work on AI.