By Matt Harrington, BCBA · Behaviorist Book Club · April 2026 · 12 min read

Augmented Intelligence and Functional Behavior Assessment: Navigating the Intersection of AI and Clinical Judgment

Overview & Clinical Significance

Functional behavior assessment sits at the heart of behavior analytic practice. Every intervention plan worth its name begins with understanding why a behavior occurs, and the FBA is the vehicle through which that understanding is developed. Now, artificial intelligence tools are entering this process, not as replacements for clinical reasoning but as augmented intelligence systems that can assist with data organization, pattern recognition, and hypothesis generation.

Adam Ventura's session on conducting FBAs with AI tackles a question that many practitioners have been quietly asking: where does AI fit in an assessment process that has traditionally depended entirely on human observation, clinical experience, and professional judgment? The answer, as this course frames it, is that AI functions best as a collaborative partner rather than an autonomous assessor. Tools like Intraverbal AI can process large volumes of indirect assessment data, identify patterns across multiple informant reports, and generate preliminary functional hypotheses that the behavior analyst can then evaluate, refine, or reject.

The clinical significance of this integration lies in what it addresses. Behavior analysts conducting FBAs often work under significant time pressure, juggling multiple cases with complex presentations. Indirect assessments alone can generate pages of interview data, rating scale results, and record reviews that require synthesis. AI tools can accelerate this synthesis phase, flagging potential response classes, identifying convergence across data sources, and surfacing patterns that a clinician working through the data linearly might miss or take considerably longer to identify.

However, this acceleration introduces new questions about validity and reliability. When an AI system generates a functional hypothesis, what is the basis for that hypothesis? How does the system weight conflicting information from different informants? Does the AI's pattern recognition align with the conceptual framework of behavior analysis, or is it identifying statistical correlations that lack functional relevance? These are questions that every practitioner considering AI integration must grapple with.

The concept of interobserver agreement takes on a novel dimension in this context. Traditionally, IOA is calculated between two human observers recording the same behavior. When one of those observers is an AI system, the comparison becomes a measure of convergence between human clinical judgment and algorithmic pattern recognition. High agreement between AI-generated hypotheses and human-generated hypotheses may indicate that the AI tool is capturing meaningful functional relationships. Low agreement does not automatically mean the AI is wrong, but it does signal the need for additional direct assessment to resolve the discrepancy.

For the field of behavior analysis, this moment represents a fork in the road. Practitioners can either engage proactively with AI tools, learning their capabilities and limitations through structured evaluation, or they can wait until these tools are adopted by adjacent professions and then find themselves responding to AI-generated reports that they had no role in developing. Proactive engagement ensures that behavior analysts remain the clinical decision-makers in the assessment process.

Background & Context

The functional behavior assessment has evolved considerably since its early formulations. What began as primarily a research methodology has become a clinical staple required by educational regulations, insurance authorization processes, and professional standards. The FBA process typically includes indirect assessments such as interviews and rating scales, descriptive assessments involving direct observation in natural settings, and in some cases experimental functional analyses that systematically manipulate environmental variables to test hypotheses about behavioral function.

Each component of this process generates data that requires clinical interpretation. Indirect assessments depend on informant accuracy and may be influenced by recall bias, social desirability, or limited behavioral vocabulary among caregivers and educators. Descriptive assessments are shaped by the observer's training, the representativeness of observation periods, and the complexity of the natural environment. Functional analyses, while providing the strongest evidence for functional relationships, are resource-intensive and not always feasible in applied settings.

Artificial intelligence enters this landscape at a time when the field is simultaneously expanding in practitioner numbers and facing increased demands for efficiency. The growth in the number of BCBAs has been accompanied by growing caseloads, particularly in agencies serving individuals with insurance-funded ABA services. The pressure to complete assessments quickly while maintaining quality creates a natural opening for tools that can assist with data processing.

Augmented intelligence, as distinct from autonomous AI, refers to systems designed to enhance human decision-making rather than replace it. This distinction is critical in clinical contexts where professional judgment, contextual knowledge, and ethical responsibility cannot be delegated to an algorithm. In the FBA context, augmented intelligence might analyze interview transcripts to identify common antecedent themes, organize ABC data into visual patterns, or generate a ranked list of possible functions based on the data entered. The behavior analyst then brings contextual knowledge, clinical experience, and direct observation to evaluate these outputs.

Intraverbal AI, referenced in Ventura's session, represents one example of a tool designed specifically for behavior analytic applications. The name itself is a nod to Skinner's verbal operant, suggesting a system that processes verbal input and generates verbal output in a way that parallels the intraverbal exchanges between a clinician and an informant. Whether such tools actually operate on behavioral principles or simply use behavioral language to describe statistical processes is an important distinction that practitioners should investigate before adopting them.

The broader healthcare context also matters. AI-assisted diagnostic tools are already in use in radiology, pathology, and mental health screening. Behavior analysis is not the first clinical discipline to face questions about AI integration, and the experiences of other fields offer cautionary tales and useful frameworks. Fields that adopted AI tools without establishing clear protocols for human oversight have encountered problems with algorithmic bias, over-reliance on automated outputs, and erosion of clinical skills.

Clinical Implications

Integrating AI into the FBA process creates ripple effects across multiple dimensions of clinical practice. The most immediate implication is that behavior analysts need to develop a new competency: the ability to critically evaluate AI-generated outputs. This is fundamentally different from evaluating the output of a supervisee or colleague because the reasoning process behind AI recommendations is not always transparent or interpretable.

When an AI tool generates a hypothesis that the function of self-injurious behavior is escape from demands, the clinician must ask several questions. What data inputs led to this hypothesis? Did the system weight caregiver interview responses more heavily than direct observation data? Was the system trained on a dataset that overrepresents certain functions, potentially biasing it toward escape hypotheses for self-injury? These questions require a level of technological literacy that most behavior analysts have not yet been trained to develop.

The implications for direct assessment are particularly important. If a behavior analyst begins the assessment process by reviewing an AI-generated preliminary hypothesis, there is a risk of confirmation bias during subsequent direct observation. Knowing that the AI has flagged escape as the likely function may subtly influence what the observer attends to, how they interpret ambiguous sequences, and when they conclude that sufficient data have been collected. Structured observation protocols with predetermined recording criteria can mitigate this risk, but only if the clinician is aware of the bias potential in the first place.

For organizations, AI integration into assessment has implications for quality assurance and peer review processes. If FBA reports increasingly incorporate AI-generated components, clinical reviewers need standards for evaluating those components. Questions such as which AI tool was used, what version of the software generated the output, and whether the clinician modified the AI-generated hypothesis based on additional data should become standard elements of report review.

The training implications extend to supervision as well. Supervisors working with trainees who have access to AI tools may find that trainees are generating more polished-looking assessment reports more quickly, but the supervisor cannot easily determine whether the trainee actually understands the conceptual basis for the functional hypotheses presented. A trainee who accepts an AI-generated escape hypothesis without understanding the three-term contingency underlying escape-maintained behavior has not actually learned to conduct an FBA. Supervisors may need to implement structured competency checks that separate the trainee's clinical reasoning from the AI's output.

There is also a practice implication around informed consent. Families and other stakeholders have a right to know how their assessment data will be processed. If interview responses are being entered into an AI system, the family should understand this before participating. This is not a minor procedural point. Families may have concerns about data privacy, algorithmic decision-making, or the role of technology in their child's assessment that deserve to be addressed directly and honestly.

Finally, the convergence between AI-generated and human-generated hypotheses, measured through IOA-like comparisons, has implications for how we validate these tools. A tool that consistently agrees with experienced clinicians may be useful for efficiency but adds little new insight. A tool that sometimes disagrees in ways that prompt the clinician to reconsider their initial hypothesis may actually improve clinical outcomes by challenging assumptions.

Ethical Considerations

The integration of artificial intelligence into functional behavior assessment raises ethical questions that touch nearly every section of the BACB Ethics Code for Behavior Analysts. At the broadest level, the obligation to practice within one's boundaries of competence (Code 1.05) requires behavior analysts to honestly assess whether they have sufficient understanding of AI tools to use them responsibly. Using an AI system without understanding its limitations, its data handling practices, or its potential for systematic error is analogous to administering an assessment instrument without understanding its psychometric properties.

Data privacy stands as one of the most pressing ethical concerns. When assessment data is entered into an AI platform, questions arise about where that data is stored, who has access to it, whether it is used to train the AI model further, and whether the data can be de-identified sufficiently to protect client confidentiality. Code 2.04 addresses the behavior analyst's responsibility to protect confidential information, and this responsibility does not diminish because a third-party technology platform is involved. Behavior analysts must conduct due diligence on the data practices of any AI tool before entering client information.

The question of clinical decision-making authority is equally significant. Code 2.01 establishes that behavior analysts provide services that are conceptually consistent with behavior analytic principles and supported by evidence. An AI-generated functional hypothesis is neither inherently consistent nor inconsistent with behavior analytic principles; it depends on the model's underlying logic. If a behavior analyst accepts an AI-generated hypothesis without independently verifying it through direct assessment, they may be delegating clinical judgment to a system that operates outside the conceptual framework of the profession. The ethics code places decision-making authority with the behavior analyst, and no AI tool alters that responsibility.

Informed consent requires particular attention in this context. Code 2.11 mandates that clients and relevant stakeholders are informed about the assessment process in a manner they can understand. If AI tools are being used to process assessment data, this should be disclosed to families. The disclosure should include what the tool does, what data it receives, and how its outputs are used in the assessment process. Families should have the opportunity to ask questions and, where feasible, to decline AI involvement without this affecting the quality of services they receive.

Supervision ethics (Code 4.0) become more complex when supervisees have access to AI tools. A supervisor who allows a trainee to use AI-assisted FBA tools without providing guidance on their appropriate use, limitations, and ethical implications is not meeting their supervisory obligations. Conversely, a supervisor who prohibits AI use entirely without explanation may be limiting a trainee's exposure to tools that are becoming part of the professional landscape.

There is also an equity consideration. AI systems are trained on datasets, and those datasets may not represent the full diversity of the populations behavior analysts serve. If an AI tool was primarily trained on data from English-speaking, middle-class families, its hypotheses may be less accurate for families from different cultural, linguistic, or socioeconomic backgrounds. Using such a tool without awareness of its training data limitations risks introducing systematic bias into the assessment process, which conflicts with the obligation to provide culturally responsive services (Code 1.07).

Assessment & Decision-Making

Deciding whether and how to integrate AI into the FBA process requires a structured decision-making framework rather than an all-or-nothing approach. The first step is evaluating the specific AI tool under consideration. Not all AI systems marketed for behavioral assessment are created equal, and behavior analysts should approach tool selection with the same rigor they would apply to selecting a standardized assessment instrument.

Key evaluation criteria include the tool's transparency about its methodology, the training data used to develop the model, any validation studies comparing its outputs to human clinical judgment, its data security practices, and its intended scope of use. A tool designed to assist with organizing indirect assessment data is a fundamentally different proposition from one that claims to generate functional hypotheses from minimal input. The former supports the clinician's workflow; the latter attempts to replicate the clinician's reasoning.

Once a tool has been evaluated and deemed appropriate, the next decision involves determining which FBA components are suitable for AI assistance. Indirect assessment data processing is a strong candidate because it involves synthesizing large volumes of verbal report data where pattern recognition is valuable. Organizing ABC data collected during descriptive assessments into visual displays and frequency summaries is another area where AI can add efficiency without displacing clinical judgment. Hypothesis generation can benefit from AI involvement, provided the output is treated as a preliminary suggestion rather than a conclusion.

Direct observation and functional analysis, by contrast, are components where AI involvement should be more limited. These assessment phases depend on real-time clinical judgment about what to observe, how to interpret ambiguous behavioral sequences, and when to adjust conditions based on emerging data. While AI might eventually assist with video-based behavioral coding, the current state of the technology is not at a point where it can replace trained human observation for most behavior analytic applications.

A practical implementation framework might proceed as follows. First, conduct the initial intake and indirect assessments using standard clinical methods. Second, enter the indirect assessment data into the AI tool and review its output, including any preliminary hypotheses, identified patterns, or flagged inconsistencies. Third, use the AI output to inform, but not determine, the direct observation plan. If the AI identifies potential antecedent patterns that the clinician had not considered, include those in the observation protocol. If the AI hypothesis conflicts with the clinician's impression, plan observations that will help resolve the discrepancy. Fourth, conduct direct assessments and, if indicated, functional analysis based on the combined clinical and AI-informed observation plan. Fifth, synthesize all data sources, including the AI output, clinical observation, and any experimental data, into a final hypothesis and intervention plan authored entirely by the behavior analyst.

This framework preserves the clinician's role as the primary decision-maker while leveraging AI for its strengths in data processing and pattern detection. It also creates a natural checkpoint where the clinician evaluates AI output before it influences subsequent assessment phases.

For organizations implementing AI tools across multiple clinicians, establishing inter-rater reliability protocols between clinicians using and not using AI assistance can help determine whether AI integration is actually improving assessment quality or simply changing its form. Comparing FBA outcomes, treatment effectiveness, and family satisfaction across AI-assisted and non-AI-assisted cases would provide the most meaningful evaluation data.

What This Means for Your Practice

If you are considering AI tools for your FBA process, start by identifying the specific bottleneck you are trying to address. If the challenge is processing large volumes of indirect assessment data across multiple informants, AI tools designed for data synthesis may offer genuine efficiency gains. If the challenge is generating functional hypotheses for complex cases, AI output should be viewed as one data point among many rather than a shortcut to clinical conclusions.

Before adopting any tool, investigate its data security practices and ensure they meet the standards required by your organization, your state licensing board, and HIPAA regulations. Read the terms of service carefully, particularly provisions about data retention, model training on user data, and third-party access. If the tool's privacy practices are opaque, that is a disqualifying factor regardless of how useful the clinical features appear.

Develop a disclosure protocol for families. Prepare a brief, plain-language explanation of what the AI tool does, what data it receives, and how you use its output. Practice delivering this explanation before you need to do it in a clinical setting. Most families will be understanding when the tool is framed as something that helps you organize information more efficiently, but they deserve the opportunity to ask questions.

If you supervise trainees, establish clear guidelines for AI use during the assessment process. Require trainees to complete FBAs without AI assistance before introducing AI tools, so that you can verify their independent clinical reasoning skills. When AI is introduced, build in supervision activities where the trainee presents their own hypothesis alongside the AI's hypothesis and explains any discrepancies.

Finally, treat your experience with AI-assisted FBAs as data. Track whether assessments completed with AI assistance take less time, whether the functional hypotheses they produce are equally or more accurate compared to your non-AI-assisted work, and whether families report understanding the assessment process. This personal outcome data will help you make informed decisions about continued AI integration rather than relying on marketing claims or peer enthusiasm.