Behavior Analysts as AI Critics: A Professional Responsibility

On a Friday morning that felt indistinguishable from most others, Maya noticed something odd.

Maya is an RBT, sharp, conscientious, and six months into a caseload that had taught her more about real-world variability than any textbook ever could. She was reviewing a session summary generated by a new “AI-assisted documentation” feature her clinic had recently enabled. The note looked clean, confident, and fluent. Too fluent.

She toggled back to her raw data. A discrepancy surfaced, small but unmistakable. The AI summary described “consistent engagement during instructional trials.” Maya’s event-based data told a different story: engagement was variable, context-sensitive, and dipped reliably after schedule changes. Nothing catastrophic, but it was not “nothing,” either.

She flagged it and walked down the hall to Alex’s office.

Alex is the supervising BCBA. Fifteen years in the field, calm under pressure, deeply allergic to hype. Alex had approved the AI tool because, on paper, it lived in what the vendor called a “low-risk efficiency zone.” It summarized notes, it didn’t make clinical decisions. At least, that was the claim.

This moment marks the beginning of the story, not because anything went wrong yet, but because this is how responsibility begins in applied behavior analysis. Not with catastrophe, but with noticing.

The Call to Critique, Not Compliance

Applied Behavior Analysis (ABA) has always been a field of critics. Not critics in the cultural sense, but in the functional sense. We interrogate claims. We trace effects to contingencies. We ask whether something works, under what conditions, and at what cost.

AI does not change that responsibility. It intensifies it.

Being “in the loop” with AI systems is often framed as oversight. Review the output. Approve the recommendation. Click confirm. But that framing is passive, and passivity is not a behavior-analytic virtue.

In Project Chiron terms, being in the loop is not about supervision, it is about critique.

Alex looked at Maya’s flagged note and said something deceptively simple:

“Let’s slow down and figure out what this thing is doing.”

That sentence contains the entire orientation shift this moment requires.

The Critic’s Toolkit: Four Skills That Matter

Alex didn’t start by disabling the tool or defending it. Instead, Alex turned the moment into supervision, not just of Maya, but of the system.

What follows is the critic’s toolkit Alex walked Maya through, one skill at a time.

1. Claim Parsing: What Is the System Actually Saying?

The vendor’s website claimed the tool “accurately summarizes session engagement patterns.”

Alex paused there.

“Accurately compared to what?” Alex asked.

“Over what time window?”

“Using which signals?”

“And what does ‘engagement’ even mean in their model?”

Claim parsing is the skill of breaking vendor language into testable components. Not debating intentions, not inferring ethics, but isolating assertions.

Maya learned to translate marketing language into behavioral questions:

• What inputs does the model use?

• What outputs does it generate?

• What dimension of behavior is it compressing?

• What variability does it discard to do that?

In other words, what is reinforced in the system’s loss function, even if no one names it that way.

The AI was not lying. It was optimizing.