If a Skill Only Shows Up in the Clinic, Did It Actually Generalize?

Here's a question ABA doesn't ask itself often enough: how do we actually know a skill generalized?

Not "did the child hit criterion on three consecutive probes." Not "did the graph trend upward." Whether the child, at home, at school, at the grocery store, on a random Tuesday with no therapist in the room, actually uses the skill the way it was intended to be used.

Most programs can't answer that question with real data. They answer it with a proxy: a generalization probe, run occasionally, in a semi-artificial setup, by someone holding a clipboard. It's not nothing. But it's a snapshot standing in for something that's supposed to be continuous — and everyone in this field knows the snapshot and the reality don't always match.

The blind spot nobody names out loud

This isn't a flaw in any particular clinician or program. It's structural. Discrete Trial Training produces excellent data precisely because it happens in a controlled setting with someone dedicated to observing and recording. That control is the whole point — and it's also the reason the data mostly describes what happens in that room, not what happens in the child's actual life.

Naturalistic teaching approaches — Incidental Teaching, Pivotal Response Treatment, Natural Environment Teaching, and the rest of the NDBI family — exist specifically to close that gap, embedding learning in real routines instead of a clinic room. The evidence behind them is solid. But here's the part that rarely gets said plainly: even naturalistic teaching, as it's practiced today, is still bounded by session time. A therapist doing NET in a child's living room for forty-five minutes is still only capturing forty-five minutes of a sixteen-hour day. What happens at 7am during breakfast, or at recess, or during a meltdown at a cousin's birthday party three weeks later, isn't in anyone's data set. It's not that the field doesn't value generalization — it's that continuous measurement of it has never actually been possible. So the field built its outcome claims on the best available proxy and mostly stopped questioning the proxy.

What changes when data collection stops being a room

This is the part that's genuinely new, not a rehash of "AI makes note-taking faster." Passive sensing — wearables tracking movement, physiological state, and behavior markers; computer vision that can code affect and joint attention without a human doing momentary time sampling by hand — doesn't just make in-session data collection easier. It removes the constraint that data collection has to happen inside a scheduled session at all.

That's a different kind of claim than convenience. It means, for the first time, a real possibility of knowing whether a skill taught on Monday actually shows up on Thursday, in a setting no therapist ever walked into. It means dosage questions — how much intervention, at what intensity, produced how much real-world change — stop being unanswerable in principle, because the outcome side of the equation is no longer limited to a handful of clinic-based probes. And it means fidelity, currently one of the most self-reported and least verifiable parts of ABA quality assurance, could be grounded in something closer to what actually happened, continuously, instead of a checklist filled out after the fact.

None of this is about replacing the therapist's judgment. It's about giving that judgment something it's never had before: a real signal from the parts of a child's life where nobody was watching with a data sheet.

Why this matters more than it sounds like it does

This reframes what "generalization" even means as a construct in the field. Right now, generalization is treated as a phase — you teach a skill, then you check whether it generalized, as a separate step, usually near the end of a goal's lifecycle. If continuous data becomes real, generalization stops being a phase you check and becomes something you can actually watch happen, or fail to happen, in close to real time. A skill that isn't transferring outside the clinic would show up as a signal within days, not get discovered months later during a periodic review.

It also changes who gets to be believed. Insurance payers have long-standing skepticism about ABA outcomes data, partly because the people collecting the data are the same people whose services are being evaluated by it. Continuous, passively generated data from outside the clinic isn't produced by anyone with an incentive to show progress — it's just what happened. That's a fundamentally different kind of evidence than a provider-generated progress report, and it's the kind of evidence that could shift how this field is trusted from the outside, not just how efficiently it operates on the inside.

The honest caveat

Continuous monitoring of a child, even for good reasons, is not a small thing to propose, and it shouldn't be pitched as an unambiguous win. Consent, data ownership, and where the boundary sits between "helpful measurement" and "surveillance of a child who didn't choose this" are real questions, not disclaimers to bury in fine print. Families deserve to know exactly what's being captured, who sees it, and how to turn it off. Any product built on this idea earns trust by treating that seriously, not by treating it as friction to design around.

But the underlying claim is worth sitting with: the field has spent decades getting extremely good at measuring what happens in forty-five-minute windows, in rooms built for measurement. It has never had a real way to know what happens the rest of the time. That's not a paperwork problem. It's the actual outcome question ABA has been answering with a proxy for as long as the field has existed — and it's the first time the proxy might not be necessary anymore.