Ethical Use of Artificial Intelligence in Behavior Analytic Practice

Overview & Clinical Significance

Artificial intelligence is entering behavior analytic practice through multiple channels simultaneously: AI-assisted documentation tools, data visualization platforms, session analysis software, and communication aids all now incorporate machine learning components. Practitioners who use these tools without understanding their ethical implications take on risks they may not be aware of — and, under the BACB Ethics Code (2022), responsibility for those risks does not transfer to the AI system.

This CEU addresses the framework developed by the Consortium for Ethical Artificial Intelligence in Applied Behavior Analysis, which identifies core ethical concerns and provides practical guidance for practitioners and organizations navigating AI adoption. The clinical significance is immediate: AI tools that introduce errors, generate biased outputs, or expose client data without appropriate safeguards directly affect client welfare — the Ethics Code's primary concern.

The framework centers on four properties: truthfulness (AI outputs should be accurate and not mislead), accountability (human practitioners remain responsible for AI-assisted decisions), transparency (clients and families should know when AI is used in their care), and client welfare (AI adoption should demonstrably benefit, not harm, the individuals served). These properties map directly onto existing Ethics Code provisions.

The research base on intervention effectiveness for individuals with autism continues to grow — including work on caregiver report accuracy in treatment evaluation (Pichardo et al. (2026)) and meta-analysis of single-case interventions (Kok et al. (2026)).

AI tools that claim to summarize or apply this literature require careful scrutiny — their outputs are only as good as the data they were trained on and the assumptions built into their design.

The Consortium's proactive stance — developing field-specific guidance before regulatory frameworks require it — is itself ethically significant. It reflects the values-based orientation of the 2022 Ethics Code: acting in clients' best interests does not wait for rules; it anticipates challenges and builds appropriate frameworks in advance. BCBAs who engage with the Consortium's guidance and participate in its ongoing development are contributing to field-level infrastructure that protects clients, supports practitioners, and positions behavior analysis as a responsible adopter of emerging technology rather than a reactive one.

Background & Context

The pace of AI adoption in healthcare settings has outrun the development of ethical frameworks for its use. Behavior analysis is not unique in this regard — medicine, psychology, and education are all navigating the same challenge. What is distinctive about ABA is the population served: individuals with autism and intellectual disabilities who may have limited ability to consent to or contest AI-assisted decisions made about them.

The Consortium for Ethical Artificial Intelligence in Applied Behavior Analysis represents an organized professional response to this challenge. Rather than waiting for regulatory frameworks to catch up with practice, behavior analysts are developing field-specific guidance that can be applied now — consistent with the values-based orientation of the 2022 Ethics Code revision.

AI tools in clinical practice raise distinct concerns depending on their function. Documentation AI raises accuracy and consent concerns. Data analysis AI raises interpretability and accountability concerns.

Communication AI raises truthfulness and competence concerns. Each function requires specific ethical scrutiny rather than a one-size-fits-all approach.

The research on precision teaching and measurement of private events (Van & Kubina (2026)) is relevant here: AI tools that attempt to infer internal states or motivation from observable behavior are doing something technically ambitious and ethically fraught. The assumptions built into such inferences may be invisible to the practitioner using the tool — and erroneous AI-generated inferences may shape clinical decisions without the practitioner recognizing the error.

Mealtime behavior research in autism (Tong et al. (2026)) illustrates how population-specific factors shape intervention design. AI tools trained on data that underrepresents particular populations will produce recommendations that are less applicable to those populations — a direct client welfare concern that practitioners must actively evaluate.

The single-case research methodology that underpins ABA's evidence base provides a particularly useful model for evaluating AI tool performance: treat the AI as an intervention, establish a baseline, implement the tool, collect data on outcomes that matter clinically, and evaluate effectiveness with the same rigor applied to any other intervention. This is not a theoretical suggestion — it is a practically implementable approach that any organization with data collection infrastructure can operationalize. Practitioners who approach AI adoption with this evaluative mindset are far better positioned to catch problems early, adjust their use of AI tools based on evidence, and discontinue tools that do not produce the intended clinical benefits.

Clinical Implications

The clinical implications of AI adoption depend substantially on which tools are used, for which purposes, and with what oversight structures in place. The Consortium's framework provides a practical starting point for each of those decisions.

For documentation AI, the key clinical implication is that the practitioner remains responsible for the accuracy of every note submitted to insurance, included in the client's record, or used to inform treatment decisions. An AI that generates an inaccurate session note or a biased behavioral summary creates liability for the practitioner who accepts it without review. The Ethics Code's provisions on truthful documentation apply regardless of whether a human or an algorithm drafted the initial text.

For data analysis AI, the clinical implication involves interpretability: if the practitioner cannot explain how the AI reached its conclusion, they cannot responsibly act on that conclusion. Research on caregiver report accuracy (Pichardo et al. (2026)) illustrates that even well-motivated informants can provide inaccurate data — AI systems require the same critical evaluation, applied with the same methodological rigor BCBAs bring to any clinical data source.

For communication AI, the implication is direct: AI outputs that appear under the practitioner's name without substantive review function as unlicensed practice. The practitioner's license is what families trust — that trust requires active, not passive, endorsement of every output associated with clinical care.

The single-case research methodology that underpins ABA's evidence base (Kok et al. (2026)) provides a model for evaluating AI tool performance: treat the AI as an intervention, measure its effects on clinically relevant outcomes, and revise or discontinue based on what the data show.

Consent for AI use should also be revisited over time. As AI tools are updated, as new features are introduced, or as an organization's use of AI expands, the original consent may no longer accurately describe what clients and families are agreeing to. Best practice involves treating consent for AI use as a living agreement that is reviewed and updated as the tools and their uses evolve — the same standard that applies to consent for any aspect of clinical practice that changes over time.

The research on mealtime behavior intervention effectiveness (Tong et al. (2026)) illustrates how behavioral and contextual factors interact in complex ways — AI tools that flatten that complexity in their outputs may produce summaries that are technically accurate but clinically misleading, making active practitioner oversight essential.

Ethical Considerations

The BACB Ethics Code (2022) was not written with AI in mind, but its provisions apply to AI adoption with full force. The responsibility-not-transferred principle is perhaps the most critical: using an AI tool does not shift ethical responsibility from the practitioner to the technology. If an AI generates a harmful recommendation and the practitioner implements it without critical review, the practitioner is accountable for the outcome.

Accountability in the Consortium's framework means that organizations and practitioners must be able to identify who decided to use an AI tool, who reviewed its outputs, and who acted on its recommendations. That accountability chain must be maintained even when AI is used at scale across hundreds of sessions.

Transparency with clients and families involves more than disclosure — it requires explanation. What the AI does, what data it uses, what errors are possible, and what the practitioner's review process looks like. Families of individuals with autism, who may have had their children subjected to practices they did not fully understand, deserve more than a checkbox on a consent form.

The client welfare property requires asking whether AI adoption actually benefits clients or primarily benefits organizations. AI tools that reduce practitioner documentation burden are clinically neutral at best — unless the time saved is reinvested in direct client contact or clinical quality. Research on mealtime behavior problems (Tong et al.

(2026)) illustrates that clinical decisions require nuanced, population-specific information — AI tools that flatten that nuance may produce recommendations that are efficient but clinically inappropriate.

The transparency property of the Consortium's framework also has implications for how AI is discussed in research and quality improvement contexts. Organizations that use AI in session analysis, treatment planning, or outcome monitoring should be transparent about that use in any publications, reports, or quality improvement documents that draw on AI-assisted data. The same honesty norms that govern clinical documentation govern research and quality improvement reporting — AI use is a methodological choice that should be disclosed, not a background process that can be assumed or omitted.

Assessment & Decision-Making

Assessing AI tools before adoption requires a structured evaluation process that maps to the Consortium's four properties. For each AI tool under consideration, practitioners and organizations should ask: Is this tool's output accurate and verifiable? Who is accountable for errors?

Are clients and families informed about its use? Does the evidence support that it benefits clients?

The truthfulness assessment requires examining the evidence base for the AI tool's performance. Vendors often provide accuracy statistics that reflect best-case performance on curated data — real-world performance, particularly with populations that differ from the training set, may be substantially lower. The methodology used in meta-analyses of single-case research (Kok et al.

(2026)) provides a template: examine effect sizes, confidence intervals, and variation across contexts before drawing conclusions about a tool's adequacy.

The accountability assessment requires mapping decision authority before adoption: who reviews AI outputs, at what frequency, using what criteria, and with what escalation path for identified errors. Those policies should be maintained prospectively, not reconstructed after a problem occurs.

For individual practitioners, decision-making about AI adoption also involves an honest self-assessment of competence: do I understand this tool well enough to use it responsibly? The Ethics Code's competence provisions apply to technological tools as much as to clinical techniques. Research on private event measurement (Van & Kubina (2026)) demonstrates that even sophisticated behavioral constructs can be operationalized and measured — practitioners should apply that same discipline to their own AI literacy before adopting tools in clinical contexts.

AI adoption decisions should also consider the organizational context: does the organization have the technical capacity to implement the AI tool appropriately? Does it have the supervision structures to maintain practitioner oversight at scale? Does it have the incident response capacity to address harms when they occur?

A tool that might be appropriate for a large, well-resourced organization with dedicated technical support may be inappropriate for a small practice without those resources. The Consortium's four-property framework applies to organizational readiness as much as to tool quality.

What This Means for Your Practice

Before adopting any AI tool for clinical use, apply the Consortium's four-property framework as a structured checklist. That evaluation should happen before implementation, not after a problem surfaces. If a tool cannot be evaluated against those criteria — because the vendor does not provide transparency about its methodology, training data, or error rates — that opacity is itself a red flag.

If AI tools are already in use in your practice, conduct a retrospective audit. Review a sample of AI-generated outputs and assess their accuracy against your independent clinical knowledge of those cases. That audit will tell you where current oversight is adequate and where it needs strengthening.

For informed consent, develop disclosure language that actually explains AI use rather than burying it in standard consent forms. Families deserve to understand how AI is used in their child's care and what safeguards are in place. That conversation may feel uncomfortable, but it is ethically required — and most families will appreciate the transparency.

Stay engaged with the Consortium and related professional resources as this area develops rapidly. Practitioners who participate in developing the ethical framework for AI in ABA — rather than waiting for finalized guidance — contribute to field-level protections that will benefit clients and practitioners alike.

For practitioners in leadership roles — clinical directors, program managers, owners — AI adoption decisions have organizational implications that individual practitioners do not face alone. Developing an organizational AI governance structure before it is needed — with clear policies, designated oversight responsibilities, and regular review processes — provides the infrastructure that makes individual practitioner accountability sustainable. The research on single-case methodology (Kok et al.

(2026)) illustrates how individual-level data aggregate into organizational-level understanding — the same aggregation logic applies to monitoring AI performance across a clinical organization.