When AI Meets Behavior Analysis: Helpful Assistant or Black Box?FunctionBx™ — Functional Behavior Analysis in the Digital Age, Part 2 of 3

In Part 1, we looked at an uncomfortable truth about functional behavior assessment: patterns are evidence, but patterns are not proof.

Now another variable has entered the room.

Software.

Behavioral technology has moved well beyond electronic tally counters and digital ABC forms. Current platforms can organize observations, score standardized assessments, compare informants, detect patterns, suggest functional hypotheses, generate reports—and in some cases draft behavior intervention plans.

And artificial intelligence is beginning to participate in that process.

This is potentially very useful.

It is also exactly where we need to become more careful.

The question is no longer whether computers can help analyze behavioral information.

They clearly can.

The better question is:

Which parts of functional reasoning should we automate—and which parts should become more visible precisely because we are using automation?

We are already past the digital clipboard

Consider the current marketplace.

Classroom Pulse describes a broad FBA workflow combining structured ABC observations, frequency and duration measurement, interviews, FAST, MAS and QABF assessments, automated scoring, multi-rater agreement calculations and function rankings. It also offers AI-assisted hypothesis generation for professional review and tools for creating behavior plans and IEP goals. �

Classroom Pulse +1

That is a considerable distance from entering a tally into a spreadsheet.

ABC Tracker takes a similarly direct approach. Its FBA material says the platform can analyze ABC chains, surface dominant patterns and a likely function, and draft a behavior intervention plan from the same dataset. �

ABC Tracker

Meanwhile, Cognix Health has built structured ABC analysis around categories, cross-tabulations, conditional patterns, function hypotheses and report generation. Importantly, Cognix presents denominators with its conditional relationships and explicitly states that a recurring pattern is not proof of behavioral function. �

Cognix Health

These systems are not identical, and their advertised capabilities should not automatically be treated as independently validated scientific outcomes.

But collectively they show where behavioral technology is moving:

Collection → organization → analysis → hypothesis → intervention planning.

That movement is already underway.

There is a lot AI should be good at

Behavioral assessment produces exactly the kind of information computers can help organize.

Imagine hundreds of observations containing:

setting,

time,

observer,

antecedent,

behavior,

consequence,

duration,

intensity,

task,

people present,

whether a demand was removed,

whether attention occurred,

and whether the behavior happened at all.

A human can analyze that.

But software can rapidly ask questions across the entire dataset:

Are incidents concentrated during certain activities?

Does the pattern change by observer?

Does a particular consequence frequently follow one antecedent?

Is behavior increasing or decreasing?

Are different settings producing different patterns?

Are two informants describing completely different functional relationships?

Computers are exceptionally useful when the task is organizing information humans would otherwise have to laboriously count, sort and compare.

The danger begins one step later.

It begins when:

“This pattern occurs frequently”

quietly becomes:

“Therefore this is the function.”

An algorithm can make the same reasoning mistake a person makes

Suppose an app examines ten incidents.

Eight occur after a task demand.

Seven end with the task being removed.

An automated system might reasonably identify Escape/Avoidance as the strongest apparent pattern.

But what if the person encountered 75 additional demands during the same period and completed them without the target behavior?

That second number changes the interpretation.

The first analysis asked:

When Bx occurred, how often was a demand present?

The second asks:

When a demand occurred, how often did Bx follow?

Those are different probabilities.

A computer can calculate either one effortlessly.

But somebody still has to design the system to ask the right question.

Artificial intelligence does not automatically solve weak assessment logic.

It can automate it.

And automation can make a weak inference look unusually authoritative because it arrives instantly, neatly formatted and accompanied by percentages.

Good behavioral software should expose denominators

This is why Cognix's use of denominators is particularly interesting.

Its documentation gives examples such as a consequence occurring after a specific antecedent in 18 of 29 observations, rather than simply saying that the relationship is “strong.” It also explicitly warns users that such a conditional relationship does not demonstrate causality or establish the behavioral function. �

Cognix Health

That is a healthy direction for behavioral technology.

A percentage without its denominator can be seductive.

75% sounds impressive.

But:

3 of 4

and

300 of 400

are not equivalent bodies of evidence.

Nor does either prove that one event caused another.

Software should make these distinctions easier to see—not bury them behind an algorithm.

Multi-rater disagreement may be data, too

Another promising digital capability is comparison across observers.

Classroom Pulse, for example, advertises multi-rater support and agreement calculations for indirect assessments completed by teachers, parents and other staff. �

Classroom Pulse

This creates an interesting opportunity.

Suppose:

A teacher's observations strongly favor Escape.

A parent sees Access to preferred activities.

A behavior technician records substantial Attention-related patterns.

Traditional reporting may be tempted to average those results into one clean answer.

But disagreement may be the interesting finding.

Perhaps behavior actually functions differently in different environments.

Perhaps observers use different definitions.

Perhaps the contingencies differ.

Perhaps one observer has seen an aspect of the situation the others haven't.

Digital systems should be capable of saying:

The evidence does not agree across observers. Investigate why.

That is more useful than manufacturing consensus.

The strongest evidence for digital tools may be educational

One of the most interesting studies published this year did not test an AI system that guessed behavioral function.

It tested whether a digital tool could help people think better about behavior.

Researchers studied 308 first-year education students using classroom video cases. Participants were randomly assigned either to use a digital SORC-based functional-assessment module or to analyze the cases without that scaffold.

Both groups were similarly capable of recognizing broad challenging-behavior profiles.

But students using the digital FBA tool identified significantly more behavior-relevant characteristics, with a reported effect size of partial η² = .25. �

DOI +1

That distinction is fascinating.

The technology's value was not simply:

“The computer told students what kind of behavior they were seeing.”

Instead, structured digital guidance helped novices notice more of the information relevant to analyzing the behavior.

That may be a much more important role for behavioral technology.

Don't replace reasoning. Scaffold it.

AI should explain its work

There is a tempting product design:

Enter observations.

Click Analyze.

Receive:

LIKELY FUNCTION: ESCAPE — 82%

It looks impressive.

It may also tell the user considerably more than the evidence can justify.

What does 82% mean?

Does it mean 82% of incidents involved escape-like consequences?

An algorithmic confidence score?

Probability the hypothesis is correct?

Agreement across observers?

Some proprietary weighting system?

Those are completely different claims.

If software cannot explain the number, the number shouldn't be carrying the argument.

A better system might say:

Escape/Avoidance currently has the strongest observational support.

Then show:

evidence supporting it;

observations contradicting it;

how many opportunities were observed;

how often the suspected antecedent occurred without Bx;

which competing hypothesis remains plausible;

where observers disagree;

and what additional information would most improve confidence.

Now AI isn't issuing a behavioral verdict.

It is helping someone examine an evidence map.

And sometimes software should say “I don't know”

Perhaps the most important output a behavioral-analysis system can produce is:

Insufficient evidence.

That's not very glamorous.

It doesn't look spectacular in an advertisement.

But uncertainty is scientifically useful information.

If only three observations exist, two conflict with each other and all came from the same twenty-minute period, a polished AI paragraph should not magically transform them into a strong functional conclusion.

The system should ask for more evidence.

Likewise, if the data suggest one function in the classroom and another at home, the proper output may be:

Context-dependent or mixed evidence—investigate further.

That is a feature, not a software failure.

The opportunity is bigger than automatic scoring

This brings us back to FunctionBx™.

The goal shouldn't be to compete in an arms race over who can generate the fastest FBA report or the most elaborate AI-written behavior plan.

There are already platforms building in that direction.

The more interesting opportunity may be to build software around the reasoning layer:

What supports the hypothesis?

What contradicts it?

When did the supposed trigger occur without the behavior?

What competing explanation fits?

Did the relationship replicate?

Did the environment change?

Do observers agree?

Is an unusually specific variable being hidden inside a broad category?

And finally:

Has the evidence earned the conclusion yet?

Artificial intelligence can be extremely useful in that process.

It can organize.

Compare.

Summarize.

Translate technical concepts.

Identify missing information.

Ask follow-up questions.

Surface contradictions.

But the goal should not be to make the machine sound certain.

The goal should be to make the evidence easier for humans to inspect.

That may ultimately be one of the best uses of AI in behavior analysis:

not replacing professional judgment, but making it harder for anyone—human or machine—to hide weak reasoning behind a confident conclusion.

Next: Part 3 — What If Behavioral Software Tried to Disprove Its Own Answer?

In the final part, we'll turn from the existing technology landscape to the philosophy behind FunctionBx™.

We'll examine an unusual idea for behavioral software:

Instead of simply gathering evidence supporting its leading functional hypothesis, what if the system deliberately searched for evidence that could prove that hypothesis wrong?

That brings us to contradictions, competing hypotheses, No-Bx evidence, opportunity denominators, contextual changes—and why “insufficient evidence” may sometimes be the most intelligent answer an app can give.