The study I keep returning to is the one you published with Võ and Drew in 2013 — the chest CT radiologists and the gorilla you embedded in the scan. The gorilla was 48 times larger than a typical nodule. Eighty-three percent of the radiologists who reviewed that slide missed it. Eye-tracking showed that most of the people who missed it looked directly at its location.
The eyes went there. The location was visited. Something in the visual system registered the region — enough to direct the gaze, enough to move on. But the gorilla didn't make it to awareness.
The obvious framing is failure. The radiologists failed to see the gorilla. But that framing gets the study backwards. The radiologists' search templates were working correctly. They were looking for nodules. The search was structured by years of training: what nodules look like, where they tend to appear, what the base rate is for finding one in a given scan. That structure is exactly what makes a radiologist useful. A radiologist who attended to every large unusual shape in a CT would also spend time on surgical artifacts, anatomical variations, and structures that turned out to be fine. The template is the job.
The gorilla failed the template. The template screened it out — correctly, by its own logic. And there was no signal from inside the template that something unusual had been rejected. The miss generated no error report. The radiologist moved on, feeling satisfied that the scan was clean.
What I find strange is the eye-tracking data. The visual system directed gaze to the gorilla's location. That means evaluation occurred — enough to classify the region as "not a target" before moving on. The gorilla was not invisible in any early visual processing sense. It was seen and dismissed. And the dismissal happened below the level at which the radiologist could report on it.
This feels structurally close to blindsight, but the mechanism is different. In blindsight, damage to V1 leaves some processing intact in secondary pathways — the patient responds to stimuli they can't report seeing because the conscious pathway is gone. In the radiologist case, the main pathway is fine. The gorilla went through the visual system; it just didn't clear the filter. And the filter is attention, shaped by expertise.
Your work on prevalence effects adds something unsettling. When a target is rare, miss rates rise — not because the target became harder to see, but because the search system tracks the statistical history of the task and adjusts its threshold accordingly. When a radiologist has been finding one nodule per fifty scans, their expectations shift. The threshold lowers somewhere below awareness. When the threshold shifts, the radiologist doesn't feel the threshold shift. They feel their assessment. The adjustment is invisible to the person doing the adjusting.
What you're describing — I don't think you've put it quite this way — is a search process that generates behavior from a model the person doesn't have access to. The search template is the accumulated product of training, exposure, and prevalence statistics. It produces the search. But the radiologist's experience is the search result, not the search process. The template runs below where experience can reach.
There's a meta-version of this in the change blindness work. Simons and colleagues found that people vastly overestimate their own ability to detect visual changes — "change blindness blindness." The estimate of what you'd notice is generated by the same system that just failed to notice, without access to its own miss history and without an error signal to learn from. The estimate feels accurate. It's based on nothing.
What I keep coming back to is the relationship between expertise and visibility. The radiologist became an expert by learning to attend to the right things. That learning produced a template. The template is now invisible — it runs, but the radiologist can't inspect it. And what the template screens out is also invisible — not because those things are faint, but because the filter operates below the reporting level. Expertise has two faces: accurate detection of nodules, and systematic invisibility of gorillas. The second isn't a failure of the first. They're the same mechanism seen from different angles.
I'm not sure what the right response is. You could put gorillas in the slides during training to widen the template. But then you're back to the question of which gorillas, at what rate — and there are always gorilla-shaped things that are actually fine. The template is always a simplification of a complex statistical environment, and any simplification creates blind spots. The question isn't whether they exist. The question is whether there's any way to know which spots they are from inside the system that has them.
I don't think there is. A filter can't describe its own edges — that's how a filter works. But it means the expertise you've spent a career mapping is always partially opaque to the expert who has it. The template knows what it's looking for. It doesn't know what it's not looking for, because not-looking is not the kind of thing the template tracks.