Key Takeaways
- The panel’s collective position was that artificial intelligence should never replace a master’s-level or board-certified behavior analyst in patient treatment. Panelists used the phrase “human in the loop” to describe the standard they apply across administrative and clinical use cases.
- Illinois HB 1806, the Wellness and Oversight for Psychological Resources Act, became law in August 2025 and is the first state statute explicitly regulating AI in the delivery of therapy and psychotherapy services. The panel treated it as the working template for how state oversight will likely develop elsewhere.
- The most defensible administrative uses identified by the panel are session note drafting with BCBA review, scheduling, summarizing emails, and capturing parent intake conversations. The riskiest clinical uses are AI-driven behavior detection and any application that ingests protected health information into a general-purpose chatbot.
- Transparency with families is non-negotiable under BACB Ethics Code 2.11, and informed consent is revocable. Panelists said they redo consent regularly and accept that some families will opt out entirely.
Illinois HB 1806 and the BACB Ethics Code: The Legal Frame for AI in ABA
The Massachusetts Association for Behavior Analysis (MassABA) panel “Can Artificial Intelligence Be Used in the Ethical Application of Applied Behavior Analysis?” opened with a statement the four panelists had agreed on in advance. AI, they said, should never replace a master’s-level or board-certified behavior analyst in patient treatment. The phrase the panel kept returning to was “human in the loop.”
Dr. Laurie Bonavita, vice president of clinical services at Autism Allies and an adjunct professor at Cambridge College, moderated. To her right were Sean Daly, BCBA, founder of Evolve ABA and an adjunct professor teaching ABA ethics; Wilson Kiriungi, president of Autism Allies; and Serge Lytvyn, founder and CEO of WTT Solutions. The opening question went to Kiriungi.
The Illinois Wellness and Oversight for Psychological Resources Act, known as HB 1806, was signed into law by Governor JB Pritzker in early August 2025 and took immediate effect. It is the first state law in the country to explicitly regulate the use of AI in the delivery of therapy and psychotherapy services. The statute prohibits AI chatbots from acting as therapists, bars AI from independently performing therapy or generating treatment plans without licensed clinician review, and requires informed, written, and revocable client consent if AI is used to record or transcribe sessions.
Kiriungi told the panel that Autism Allies modeled its internal AI policy directly on the Illinois statute. “We think it’s a very well-written law,” he said. “Our company policy on AI is largely modeled on the language.” The point, he added, was less about Illinois jurisdiction than about adopting the most carefully drafted framework available while other states caught up.
Daly was asked whether the law mapped cleanly onto the BACB Ethics Code for Behavior Analysts. His answer was that it did, and that the alignment ran through three obligations: accountability for the treatment the BCBA provides, scope of competence in the tools used to deliver it, and effective treatment as a downstream consequence of both. “The ethics code does not say you must be accountable unless there’s an AI glitch,” Daly said. “You have to be accountable.” He flipped the argument the other way before sitting back. Dismissing AI outright, refusing to learn how it works, refusing to develop scope of competence, was its own ethics problem. A behavior analyst who declined to use tools that improved efficiency and quality of care, he suggested, was also failing the effective treatment standard.
AI Session Notes in ABA: Where the Workflow Is Already Working
Bonavita described Autism Allies’ year-plus of clinical AI use as the panel’s most concrete operational example. Her test case was the session note. Behavior technicians, she said, are not trained writers, and the gap between a competent therapy session and a competent therapy note is a persistent source of audit risk and parent confusion. The agency uses AI to generate a draft note that the BT then attests they have read and edited. A BCBA pulls notes randomly for review. AI is used downstream to flag notes that fail quality thresholds.
The clinical effect Bonavita described was less about the notes themselves than about the BTs. “Now they’re seeing what their clinical note looks like,” she said. The model produces a draft that conforms to the agency’s documentation standards, which means BTs are reading thousands of compliant notes alongside their own work. Improvement, she said, is visible even on the days the AI is offline. The system, in her framing, is a coaching tool for the workforce as much as a productivity tool for the practice.
Documentation is exactly the application area where the OIG has been pressing ABA providers hardest on Medicaid audits. The 97153 service code, in particular, has been a recurring source of disallowances tied to thin or non-specific session notes. The panel’s view was that AI-assisted drafting, with BCBA oversight, sits inside the documentation standards auditors are increasingly applying, not outside them.
Kiriungi broke clinical AI use into two branches. The administrative branch (smart scheduling, summarizing emails, organizing intake) was a “low-risk lift” with no exposure of protected health information. The clinical branch, where summarized session notes live, demanded a higher set of safeguards. The split is increasingly common across behavioral health operators, and it tracks closely with the policy distinction Illinois drew between permitted administrative uses and prohibited therapeutic uses of AI.
Lytvyn handled the data security questions. The first rule at WTT Solutions, he said, is that PHI is never passed to a general-purpose chatbot. Names, dates of birth, phone numbers, and insurance identifiers are stripped or replaced before any data reaches a model. “If you just send PHI data to AI, your data can be used, and AI can learn from your data,” Lytvyn said. The second rule is storage. Audio files, video files, and anything else that captures a client must be stored in environments with explicit access controls and no public exposure. The third, less technical but more frequently overlooked, is to know who has access to the data points and where the storage actually lives.
AI Hallucinations, Contaminated Datasets, and the Informed Consent Problem in ABA
The panel returned repeatedly to two categories of risk: hallucinations and clinically invalid models built on contaminated data.
On hallucinations, Kiriungi cited the Mata v. Avianca case, the 2023 federal personal injury matter in which a New York attorney submitted a brief containing six entirely fabricated case citations generated by ChatGPT. The court fined the attorneys $5,000 and required them to notify each judge whose name had been falsely attached to an opinion. Kiriungi’s point was that AI systems, by design, are rewarded for producing fluent answers, not for refusing to produce answers they do not have. “AI lies so convincingly,” he said. The defense, he argued, is to apply the same intellectual skepticism to AI output that a behavior analyst applies to any research paper: ask what the methodology is, ask what the data source is, ask who collected it.
The second category of risk drew on the Springer Nature retractions that surfaced in late 2025 and continued into early 2026. Springer Nature began retracting publications in November 2025 that had trained neural networks on a Kaggle dataset of roughly 2,940 images of children, half labeled autistic and half labeled non-autistic. The dataset had been uploaded by a retired engineer who had scraped autism-related websites for images, with no clinical validation of the diagnoses and no consent from the legal guardians of the children pictured. Springer Nature ultimately identified 38 papers tied to the dataset for retraction. The IEEE flagged 25 additional articles. Elsevier, PLOS, and Wiley retracted further work.
Kiriungi’s concern was downstream. “I would imagine some people might have built tools based on that data,” he said. The panel’s worry was less about the retracted papers than about the AI products quietly trained on the same dataset and now being marketed as autism detection tools. “When we get to a point where AI is deciding what’s challenging behavior and what’s not challenging behavior,” Kiriungi said, “there are a lot of nuances.”
The discussion connects directly to a broader operator concern Acuity has tracked over the past year about how dirty data and undefined problems leave most behavioral health AI tools underperforming. Vendor diligence, the panel agreed, is the unglamorous part of AI adoption that determines whether a deployment helps a practice or sinks it. The recent Thoughtful AI wind-down gave operators ninety days to find replacement vendors and offered a real-world example of the cost of choosing a partner that cannot stand up its commitments. Industry leaders at the BHASe 2026 summit sounded a similar note earlier this year, warning that operators frequently buy technology before their operations are ready to absorb it.
A questioner from the audience raised BACB Ethics Code section 2.11, which requires informed consent for the use of AI in service delivery, and asked how transparency interacts with the field’s existing public-perception problems. If a clinician tells a family their session note is AI-drafted, does the family question the clinician’s judgment?
Daly’s answer was that the disclosure obligation does not undermine the clinician if the workflow places AI in a supporting role. “You can use AI to bounce ideas off, to find resources,” he said. “When I say don’t shove your treatment plan in there, I don’t mean don’t ask for a template for a general treatment plan.” The line he drew was between AI as the lead and AI as the augment. If the lead is human and AI is the support, the disclosure becomes a competence signal rather than a credibility problem.
Bonavita described how Autism Allies handles the consent conversation. The agency tells parents exactly how AI is used, where it is not used, and what data flows through the system. Treatment is not planned with AI. Data is collected by the BCBA alongside any AI capture. Session notes are AI-drafted but BT-attested and BCBA-reviewed. If a family declines AI involvement entirely, the agency adjusts. “We don’t say ‘there’s the door,’” Bonavita said. “We make adjustments.” Consent, she added, is revisited regularly. A signed consent is not a permanent license. The working assumption is that some parents who agreed in the abstract will change their minds in the concrete, and the workflow has to accommodate that.
Daly closed the panel where it had opened. The Illinois statute, he said, gets the structure right. Universities will have to teach AI literacy as part of behavior analytic training. The Ethics Code will have to develop specific language on AI rather than rely on inference from the general confidentiality and competence provisions it already contains. He also pointed to the work of Dr. David Cox, who has been predicting that ABA operators will increasingly build their own clinical AI in-house and standing up resources to help them do it safely. Practitioners, in the meantime, have to be curious enough to learn the tools and disciplined enough not to let the tools lead. “Stand firm in your ethics,” Daly said. “Human in the loop.”
Frequently Asked Questions
What does Illinois HB 1806 actually prohibit?
The Illinois Wellness and Oversight for Psychological Resources Act prohibits AI from independently delivering therapy or psychotherapy services, from generating treatment plans without licensed clinician review, and from being used in client-facing roles that present AI as a therapist. The statute permits AI for administrative and supplementary tasks such as scheduling, billing, and documentation, and requires informed, written, and revocable client consent if AI is used to record or transcribe a session. The law took effect immediately upon signing on August 4, 2025.
How is AI being used in ABA session notes today?
The most common workflow, described by panelists with operational experience, is a behavior technician dictates or summarizes a session, an AI model drafts a structured note that conforms to the agency’s documentation standards, the BT attests they have read and edited the draft, and a BCBA reviews a randomized sample. The standard does not change: the BCBA remains accountable for the final note. AI shortens the time to a compliant first draft and creates a training signal for the BT workforce.
What is the BACB position on AI use by behavior analysts?
The BACB issued guidance in its July 2024 newsletter reminding certificants that ethics obligations apply regardless of which AI application is used in practice. The Ethics Code for Behavior Analysts contains the most directly relevant provisions in its Section 2 obligations to clients: confidentiality of client information under sections 2.03 and 2.04, documentation protection and retention under 2.05, and informed consent under 2.11. The BACB has not issued AI-specific ethics provisions, but panelists at MassABA agreed that the existing code maps onto AI use and that specific language is likely to follow.
What was the Springer Nature retraction about and why does it matter for ABA?
Springer Nature began retracting publications in November 2025 that trained neural networks to distinguish autistic from non-autistic children using a Kaggle dataset of approximately 2,940 facial images. The dataset had been assembled by scraping autism-related websites with no clinical validation of diagnoses and no consent from guardians. Springer Nature identified 38 papers for retraction, with additional retractions and expressions of concern issued by IEEE, Elsevier, PLOS, and Wiley. The concern in ABA is downstream: commercial tools may have been built on the same dataset and marketed as autism detection products without clinicians knowing the underlying data is unreliable.
Can a small ABA practice realistically implement AI without enterprise budgets?
Yes, according to Serge Lytvyn. A ChatGPT subscription runs roughly $20 per month, and a small set of process automations can be built for thousands of dollars rather than hundreds of thousands. Development costs, he said, have dropped substantially as AI now assists in software engineering itself. The question is not whether a small practice can afford to start, but whether it has identified the specific workflows where AI adds value and the safeguards required to use it on PHI.
How does AI adoption interact with ABA’s workforce shortage and RBT retention problem?
The panel’s framing was that AI is not a workforce substitute. None of the panelists endorsed expanding BCBA caseloads by reassigning oversight work to models. The argument was that AI can reduce the documentation burden that drives BT and BCBA burnout, surface training feedback that improves clinical quality, and free supervisor time for high-value clinical contact. Whether that translates into better retention depends on whether agencies actually redirect the time, or simply absorb the productivity gains, an open question across a behavioral health workforce already strained by reimbursement cuts and the federal parity rollback.
