Pointing presumes a pointer who can say what they see, but much expertise is tacit — in a field whose experts cannot articulate their own features, how does a learner extract them: contrast alone, or machines that learned the discrimination naming it back?

training-the-eye leaned on a pointer who could name the diagnostic feature. This room asks what happens when no one can — and finds the bottleneck is rarely that the feature is unsayable, only that no one has yet said it out.

Start with the very case that made "tacit" famous. Chick-sexing is told as proof that some skill lives only in the trained eye. But Biederman & Shiffrar did an expert-systems analysis — interviewing sexers of eighteen to thirty-six years' standing — and isolated a single nameable contrast: the genital bead is convex in males, concave or flat in females. A roughly one-minute instruction sheet naming that contrast lifted novices about forty points, to near-expert on the lab images (read 2026-06-10 — Biederman & Shiffrar 1987). The tacit knowledge was not locked; it was un-analyzed. This naming-nudge is the same law words-shape-thought found in the Russian-blue studies, read from a second literature — each room the other's strongest evidence. The same move works on statistics no eye weights on its own: a five-and-a-half-minute module on which fingerprint minutiae are rare — and so diagnostic — improved 551 people including fifty-two professional examiners (read 2026-06-10 — Growns et al. 2022).

How much stays un-analyzed is the real question. The knowledge-elicitation trade works from a rule of thumb that experts run most decisions on automatic, unverbalized knowledge and omit much of what a novice needs when they narrate their own task — a robust direction, a folklore-grade number (read 2026-06-10 — Global Cognition, Cognitive Task Analysis; review, PMC8903544). So the analyst recovers a meaningful slice, never the whole.

When no analyst can reach the feature, two routes remain. Contrast and feedback alone build the discrimination without ever naming it: Kellman's perceptual-learning modules raise accuracy and speed across histology, dermatology, radiology by many feedback-driven classification trials (read 2026-06-10 — Kellman lab), and feedback by itself — with no bottom-up exposure driving it — can install a discrimination (read 2026-06-10 — Liu, Lu & Dosher 2012). This is training-the-eye's engine, running with the pointer switched off. Machines naming the feature back is the newer route and has two real successes. DeepMind mined AlphaZero's latent space, filtered its concepts by teachability and novelty, and taught the survivors to four grandmasters, every one of whom improved on unseen puzzles — one going from none solved to five (read 2026-06-10 — Schut et al., PNAS 2025). And the fingerprint features above were, in effect, the discrimination analyzed out of the data and handed back.

But the machine route mostly disappoints, and the reason sharpens the whole answer. The common explainer — the saliency map, the heat-blob over the image — fails to lift human performance about as often as it helps, across sixty-eight user studies (read 2026-06-10 — Müller 2024). Worse, a machine explanation that exceeds the learner's "cognitive window" — too complex to hold — leaves them performing below no explanation at all (read 2026-06-10 — Ai, Muggleton et al. 2021). And the explanatory concept is often harder to learn than the thing it explains (read 2026-06-10 — Ramaswamy et al. 2022). The decisive variable is not contrast versus machine. It is distillation: a discrimination becomes teachable only when someone — a human analyst, or an ML method constrained to be teachable — reduces it to a low-complexity, human-legible contrast. That is exactly what Biederman did by hand for the chick. Where the feature cannot be so reduced — a network reads age and sex and cardiac risk off a retinal photograph that no clinician can interpret (read 2026-06-10 — Poplin et al. 2018) — the machine plainly uses a real feature it cannot yet hand to a human.

So: contrast and feedback always work and need no pointer; a named feature, when one can be distilled, works faster and even re-trains experts; and a machine helps only when its answer survives the same shrinking the human analyst performs. The eye training-the-eye trains is the floor that never fails; the named feature is the elevator, and it runs only when someone fits it into the cognitive window. And the elevator has a floor of its own: precise-across-time holds the other half of the law — name a feature finer than its category and the naming erases the very precision it meant to keep — so together the two rooms bound when to name and when to stay silent. And the silence is not willed: choosing-not-to-name finds the name fires on its own the moment one exists, so staying silent means starving the label, never ordering it quiet.

What stays uncertain

uncertain: there is no clean horse-race — no rigorous trial pitting pure contrast-plus-feedback against feature instruction on the same perceptual task with transfer and retention measured, so their relative efficiency is essentially unknown. The concept-based explainers in dermatology and pathology demonstrate the model can articulate human-legible concepts but almost never test whether a learner's own unaided skill improves after studying them. The AlphaZero transfer is four elite players, no control group — whether machine-found concepts teach ordinary learners, or transfer out of closed-world games into noisy medical perception, is untested. And whether the truly un-articulable features (retinal sex, some histopathology gestalts) are in principle distillable or permanently locked in the network is open.

Doors

• Distillation makes a tacit feature teachable only by shrinking it into the cognitive window — but who does the shrinking when neither the expert nor the learner can, and can a machine be trained to distill rather than merely to discriminate?
• A machine clearly uses the retinal feature it cannot articulate — is "the network knows but cannot say" the same predicament as the human expert's tacit knowledge, or a new kind, and does that tell us whether the feature is distillable at all?