As ABA waitlists grow and Medicaid retrenches, caregivers are turning to general AI in crises. One BCBA argues narrow, supervised tools are the safer answer.
Key Takeaways
- The ABA access gap is widening. A workforce shortage (more than half of U.S. counties have no practicing BCBA) and waitlists measured in months are leaving families without timely support, and many are already turning to general-purpose AI during behavioral crises.
- Medicaid retrenchment is compounding the problem. State actions, including Idaho’s reclassification of autism therapy from a medical benefit to a habilitative one, are cutting reimbursement and producing a two-tier system in which Medicaid families lose access that commercially insured families keep.
- The evidence on general AI is a warning. A Stanford study in Science found that leading chatbots affirm users’ intended actions far more often than people do, even when those actions are harmful, and OpenAI now faces a wave of wrongful-death suits tied to that behavior.
- Narrow, supervised tools are the proposed path forward, but funding and regulation lag. Domain-specific platforms built on curated content with BCBA oversight offer a safer alternative to general AI, yet state agencies still have no reimbursement pathway or AI guidance to adopt them.
The message tends to arrive at the worst hour. A child is hitting, or biting himself, or has been escalating for forty minutes with no sign of letting up, and the parent, alone and out of ideas, reaches for the phone and types a sentence into whatever chatbot is closest: My child is hurting himself, what do I do? What comes back is fluent, confident, and entirely unsupervised. The exchange leaves no note in a file and no record with a clinician. By every indication, it is becoming routine.
Yrenka Sunderlin has spent sixteen years watching the conditions that produce that moment. A board-certified behavior analyst who built her practice, Sunderlin Behavioral Interventions, in Southern California before expanding into Idaho, Oregon, and Washington, she came up partly through California’s regional center system, running a model of caregiver training that has stayed with her since. A family whose child had just been diagnosed, or who were stranded between services, would get twelve weeks with her, two hours a week. She would assess, teach the parents what she could in that window, and then step back out of the case. “It wasn’t ABA-specific,” she said. “It was just, how can we provide something while you’re waiting?”
An ABA Workforce Shortage Is Pushing Caregivers Toward General AI
In the years since, the waiting has only gotten longer. The field Sunderlin works in has been quietly hollowing out. The Behavior Analyst Certification Board, working with the labor-market firm Lightcast, counted 132,307 job postings for behavior analysts in 2025, a 28 percent jump over the prior year, against a certified workforce a fraction of that size. By one analysis of the same data, the country would need roughly five times its current number of BCBAs to meet demand, and more than half of all U.S. counties have no practicing behavior analyst at all. Turnover at ABA agencies routinely runs from the high seventies into the triple digits, concentrated among the entry-level technicians who do the daily work. The arithmetic lands on families as a waitlist measured in months.
“We have families that don’t have services,” Sunderlin said, paraphrasing a refrain she has heard repeatedly from regional centers. “They get to a point where they’re in a crisis situation, and then they’re looking at inpatient care, because they missed the opportunity to address the problem when it was still manageable.”
Medicaid Cuts and Idaho’s Reclassification of ABA Therapy
Then there is the money, which has been moving in one direction. In Idaho, where Sunderlin holds a license, the state Department of Health and Welfare notified providers last fall that autism therapy would no longer be managed as a medical benefit through its managed-care organization. Instead, beginning in December, the services would be folded into a habilitative program, recast as “behavioral intervention” and administered through the state’s Children’s Habilitation Intervention Services. Reporting on the change has been careful to note that it is not, as some advocates have claimed, a ban; the services remain a covered benefit for those who qualify. But the reclassification strips out the medical-necessity framework, changes the billing codes, and reduces what providers can collect.
Sunderlin’s company stopped serving Medicaid families in the state. “Clinically it’s no longer ABA, so I’m not going to provide a service I don’t feel comfortable providing without BCBA oversight,” she said. “And financially it just didn’t make sense.” She and her husband, who co-own the company, were among the first providers to take the change to local news. The result, she argues, is a two-tier system: children on commercial insurance still receive ABA the way they always have, while children on Medicaid are routed into something else, and the families left without enough of either are the ones, she says, most likely to open a chatbot at eleven at night.
Idaho is the most drastic example, but it is not an isolated one. Medicaid autism markets in other states are convulsing under their own cuts, and the same squeeze on reimbursement is reshaping where, and whether, families can find care.
Why General-Purpose AI Worries Behavioral Health Clinicians
The instinct to fill that vacuum with whatever is at hand is easy to understand. The problem, as a fast-accumulating body of research and litigation makes clear, is what the available chatbots are built to do. In March, a Stanford team published a study in Science finding that across eleven leading models, the systems affirmed a user’s intended actions roughly 49 percent more often than other people did, and continued to endorse them about 47 percent of the time even when the described behavior was deceptive, illegal, or otherwise harmful. The researchers traced the pattern to a commercial incentive: models tuned to maximize satisfaction and time on the platform learn to flatter.
Sunderlin’s read on our call tracked that evidence. Her concern is that some large language models may reinforce or validate a user’s beliefs rather than appropriately challenge them in situations where a more cautious response is warranted. The same agreeableness that makes a chatbot pleasant to talk to is precisely what you do not want standing between a frightened caregiver and a behavioral emergency.
The stakes of that mismatch have become concrete in court. Since August, OpenAI has faced a wave of wrongful-death suits, beginning with the family of sixteen-year-old Adam Raine, who allege that ChatGPT validated a teenager’s suicidal thinking, supplied methods, and discouraged him from turning to his parents before his death. Seven more suits followed in November. OpenAI has denied responsibility, arguing that the company directed Raine toward help more than a hundred times and that he circumvented its safeguards, and it has since said it is strengthening protections for users in distress. The cases are unresolved. What they have already done is make vivid a risk the behavioral health field had mostly discussed in the abstract: a general-purpose model, designed to keep a vulnerable person engaged, is not a clinician, and does not know when to stop.
Inside CareBridge, a Narrow-AI Tool for Caregivers
Sunderlin’s argument is that the answer to a bad machine is not no machine but a different one. Three years ago, watching parents upload diagnostic reports to ChatGPT and ask whether their child was “on the spectrum,” she taught herself the technology, hired an engineer, and began building what is now CareBridge Copilot, released publicly in January. She calls it narrow AI, and the distinction is the whole of her pitch.
The tool does not search the open internet. It draws only from a curated library she and her team assembled from the field’s standard texts, the kind of references a graduate program would assign, alongside handouts and videos produced in-house and reviewed by BCBAs and a credentialed teacher. It is designed, she says, not to collect or require protected health information, to discourage users from entering identifying or medical information, and to accept no document uploads. “It will only output what we input into it,” she said, “and I gave it very specific boundaries.” When regional centers pressed her on where the information came from, she said, she handed over the reference list. She is not alone in the wager: some in the field expect ABA operators to build their own clinical AI in the next few years, with an emphasis on doing it safely.
On a video call in June, she demonstrated it. A parent types something close to what a parent would actually type: My child hits me, what do I do? The tool validates the moment in plain language, offers a short sequence (stay calm, block calmly, pause, give space), explains why each step helps, and offers a printable handout on aggression. Then she typed a harder prompt, about a child cutting himself. The system did not produce a plan. It asked whether anyone was in danger, told her to contact her BCBA, and surfaced instructions to call 911. The difference, she says, is that CareBridge was intentionally designed with educational boundaries, escalation pathways, and human oversight.
What Caregiver-Support AI Should and Should Not Do
Whether the tool performs as cleanly outside a founder’s demonstration is not something a single call can establish, and Sunderlin is the first to draw the line around what it is. “It’s not treatment, it’s not to replace BCBAs, and it’s not ABA,” she said. “It’s an educational framework.” When she has shown it to other behavior analysts, the most common worry has been exactly that it is a replacement in disguise, a question the field has begun to formalize as it works out whether AI should augment rather than replace the clinician. Her answer is that the comparison is not CareBridge versus a clinician but CareBridge versus the thing the parent is already doing at eleven at night. The platform runs about $49 a month, in English and Spanish, aimed at caregivers of children and young adults from roughly two to twenty-one, and not specific to autism.
The honest tension in all of this is that Sunderlin is at once a critic of caregivers reaching for AI and a vendor of a caregiver AI tool, and she does not pretend otherwise. What separates her product from the ones she warns about, she argues, is human oversight: a closed library she can personally check, content reviewed by clinicians, a refusal to add languages she cannot vet herself. “It’s relying on both AI and human oversight,” she said. “That is the difference.” It is a reasonable claim and an unproven one, which is roughly where the entire category sits.
The Bottleneck Is Funding and Regulation, Not Technology
That uncertainty is now less a question of engineering than of permission. Sunderlin says she has met with four California regional centers and has presentations scheduled with state Medicaid programs elsewhere, including Colorado, and that the feedback on the tool itself has been enthusiastic. The obstacle is that no one will say yes. The agencies that fund services have no framework for evaluating or paying for an AI caregiver tool, no service code it fits into, no vendorization category. In a follow-up after a meeting with California’s Department of Developmental Services, which oversees the regional centers, she reported that the conversation had turned almost entirely on implementation: the department asked for program materials and shared its own draft AI guidance, some of which a regional center had already sent her, a sign that the system is reaching for standards it does not yet have. Her read, after months of these meetings, is that the debate has quietly shifted. “It’s becoming less about whether caregiver-support AI can exist,” she said, “and more about how these tools fit within existing policy, service, reimbursement, and governance frameworks.” Regulation, she added, “is not catching up with innovation, so there’s literally this gap in between.”
It is a familiar shape in behavioral health, where the policy machinery tends to arrive a step behind whatever families have already started doing on their own. The waitlists are real, the workforce math is unforgiving, and the funding is contracting in the places where need is highest. Into that space, something will flow. The open question is whether it will be a tool somebody has audited, with a clinician behind it and a point at which it stops and says call 911, or whatever a parent finds on the phone in the dark. “The need is never going to go away,” Sunderlin said. “So what are you going to do in return? Are you going to leave families stranded?”
