Shelby Dorsey built her own accommodation with AI. Her account points at an ABA retention problem nobody is measuring: the written-output filter.
Key Takeaways
- The filter is written output, not clinical skill: Shelby Dorsey describes apraxia and aphasia that make her writing read as disorganized to a reviewer applying a rubric. She was screened out at that step, repeatedly, before anyone evaluated the work underneath.
- She built the accommodation herself: Dorsey trained a model on her own recorded speech and writing so it could translate her work for a given audience without flattening her voice. She had never presented publicly before doing it.
- The retention implication is unspoken and expensive: ABA promotes clinicians into roles that require polished written output, and the field has no accommodation pathway at that transition. Every departure at that point is a trained clinician lost to a formatting problem.
- She draws the line at inputs, not outputs: Dorsey wants clinicians feeding the system their own ideas and vetted literature rather than asking it to generate both. She is openly uneasy about colleagues uploading treatment plans and insurance information.
Shelby Dorsey waited for ChatGPT the way she once waited outside a movie theater for a Twilight premiere, which is her own comparison and, she conceded onstage in Boston this month, a little embarrassing. She and her husband work at computers next to each other. When the thing arrived, they poked at it like a video game: what can this do, what can this do.
What it did, eventually, was solve a problem she had been carrying her entire career without a name for it that anyone else accepted. Dorsey is a Board Certified Behavior Analyst. She is also autistic and has ADHD, and she describes apraxia and aphasia alongside them. “It basically means like you can’t find the word,” she said. The practical consequence is that her writing and speech can read as disorganized to what she called the average listener, the average consumer, without any corresponding disorganization in her thinking.
That distinction is invisible to a rubric. “On paper, what I would produce would look disorganized or missing elements that a rubric would catch, and it would disqualify me,” she said, “and I didn’t have a mentor or support that was understanding that it was part of the way that my brain organizes information and words.” Before she found a workaround, she had never presented at a conference.
How She Trained AI as an Accessibility Tool, Not a Ghostwriter
The workaround was not a prompt. Dorsey trained the model on herself: her disabilities, her own recorded voice, her own writing. The function she was after was translation rather than composition, matching her work to the audience she was trying to reach “without losing my voice.” She used it to prepare her first presentation. Within a year she was speaking across the United States and fielding keynote inquiries.
Her approach sits at the conservative end of a field moving quickly. Behavior analysts in Massachusetts have worked through where AI belongs in a clinical workflow and where the ethics code stops it, and behavior analysts are building caregiver-facing tools to fill the access gap while families wait for services. Dorsey’s use is narrower than either: the output is her own work, translated.
She uses it in real time as well. Planning this very panel, she said, she would type what she was trying to convey and let the model hand it back to her in a form she could say out loud. “I need to get this across,” she would type, and it would return a sentence she could use to participate in her own conversation. Her framing of what changed is worth quoting at length rather than paraphrasing: “I was not able to live and to work in the professional space with what I felt was my full intellectual capacity, my full advocacy, because my communication had barriers, and AI has helped me cross those barriers, hurdle over them.”
She is emphatic that this is an access question, not a competence one. “Not because I don’t know how to do it. I’m an expert. I know how to do it, and I’m good at what I do,” she said. “But I’m not always good at communicating appropriately at what comes next.”
The clinical version applies that same move to her caseload. When her own daughter could not learn to tie her shoes and none of the procedures that had worked for years were landing, Dorsey fed in the behavior skills training literature and everything she had already tried, then asked what she was missing. Before, she said, that meant working through a binder of printed and tabbed articles, reading a passage aloud to her husband, pausing, moving to the next one. “In an ADHD household, that process wastes so much time,” she said, “and my kid still doesn’t know how to tie her shoe.” What she describes wanting is not a source of answers but something that can see her blind spots, which is a narrower and more defensible claim than most vendors make on her behalf.
Where ABA Loses Neurodivergent Clinicians: The Written-Output Gate
The workforce implication turns this from a personal story into a business one. ABA runs on written output that gets progressively heavier as a clinician advances: treatment plans, supervision documentation, reauthorization narratives, parent-facing summaries, and, increasingly, notes written to survive an audit. Federal auditors have made documentation quality the enforcement frontier, which raises the bar on written precision for everyone in a clinical role. A field that filters at the writing step will filter out clinicians whose disabilities live at exactly that step, and it will do so without ever recording a reason.
The workforce absorbing that loss is already stretched. Certification is at a record high, with 253,397 active Registered Behavior Technicians and 83,586 certified Board Certified Behavior Analysts as of April 1, 2026, and providers still describe scarcity, because the constraint sits in who leaves rather than who enters. a qualitative study of technicians in Florida published in Behavior Analysis in Practice grouped the reasons technicians go under four headings, among them a role that reads as transient rather than as a career. None of those headings captures the attrition Dorsey describes: a clinician who can do the job, leaving because she cannot clear the paperwork gate on the way up. Nobody counts that departure, which is the first problem with it.
The market is not absorbing losses gracefully either: Acuity has counted eleven workforce reductions across ABA, mental health, and addiction care this year. Retaining a certified clinician is cheaper than every alternative on that list.
Dorsey’s broader argument, made in another session at the same conference, is that the research already points at the remedy: reduce response effort on redundant tasks that are minimal to the job but must be done. She uses the technology for exactly that, and describes it in behavior-analytic terms. It clears executive dysfunction. It chunks a task list when she asks it to. It absorbs the part of the work that was never the work.
The paperwork load is also growing rather than shrinking. State audits have pushed documentation standards up sharply, from six-figure recoupments in Massachusetts tied to supervision records to Indiana rebuilding credentialing requirements by bulletin. Meanwhile the certification pipeline is narrowing at the top: BACB data show the gap between demand and supply widening. Any lever that keeps a competent clinician in a clinical role is worth more now than it was three years ago, and this is one nobody has costed.
The Counterargument: Does the Tool Standardize the People Using It?
The session’s moderator, Ellie Kazemi, extended the point in a direction Dorsey did not, and it complicates the optimism usefully. Kazemi, who did not begin with English as her only language, described revising her own thinking to fit the abstracts and formats the profession rewards, and said she felt the technology producing a uniformity across everyone using it. Her worry is that a tool which helps individuals clear a gate also standardizes what comes through it.
Both things can be true, and either way operators land in the same place. If the written gate is where the field loses people, examine the gate. Reviewing whether a rubric measures clinical reasoning or measures prose is cheaper than replacing a Board Certified Behavior Analyst, and it does not require buying anything.
The Limits She Sets on AI Use, and Why ABA Operators Should Note Them
Dorsey is not a maximalist about any of this, which makes her a more useful source than an evangelist would be. She said plainly that she worries about clinicians uploading entire treatment plans, insurance information, and other material she does not want in these systems, a concern that lands squarely on any provider without a written policy about what staff may paste into a consumer chatbot.
Her methodological rule is about direction of travel. “Use your ideas instead of asking it to create ideas for you,” she said. “Give it your ideas and ask it to create what you need to finish whatever it is your work.” She feeds it vetted literature so she can be confident about what it is drawing on. And she keeps the category boundary hard: if a mentor with five hours to review her work is available, she takes the mentor. “The human aspect of that is totally irreplaceable.”
For operators, the takeaway is not that AI is an accommodation policy. It is that Dorsey built her own accommodation because none was offered, and the field kept her only because she is stubborn and technically curious. The clinicians who were not are already gone. “It’s become our responsibility,” she said of the people in the room, “and we’re shaking the future.”
