Use it or lose it.

In the early 2000s, the United States went to refurbish one of its nuclear warheads and discovered something genuinely alarming: it had forgotten how to make one of the components. A material codenamed FOGBANK had been manufactured in the 1980s, then the line was shut. By the time it was needed again, the engineers who knew how to produce it had retired or died, and the process had never been documented in a way anyone could actually re-run. Rebuilding it delayed the programme by a year and ran roughly $69 million over budget — and the first attempts failed because of a subtle impurity in the original process that nobody had thought to write down (Stober, 2009).

The instructions were in the archive. The process knowledge was in people. And the people were gone.

Three layers, and only one of them is fragile

Dan Wang's distinction is the cleanest lens for this. Technology lives in three layers: tools (the machine), instructions (the recipe, the SOP, the code) and process knowledge — the tacit, embodied, hard-won know-how that lives in people and routines and cannot be fully written down (Wang, 2018, 2025).

AI is extraordinary at the first two layers. The problem is the third: process knowledge lives in people and practice, and is transmitted only by doing, repeatedly, with feedback. AI cannot store it — only its shadow. FOGBANK lost the third layer while keeping the first two. That is the failure mode AI now quietly invites on every shop floor — not by erasing knowledge, but by removing the conditions under which it is made.

The mechanism: capacity-hostile environments

Deskilling is not a personal weakness — it is structural. AI creates what Ferdman (2025) calls "capacity-hostile environments": settings where the design of the tool actively impedes the cultivation of the human capacities it depends on. Tacit knowledge is built by doing — the socialisation step in Nonaka's SECI model, learned at the elbow of someone more experienced. Remove the doing and you do not transfer the knowledge to the machine. You simply stop making it.

And the effect is measurable. In a 2025 study in The Lancet Gastroenterology & Hepatology, experienced endoscopists who had grown used to AI flagging pre-cancerous growths saw their unassisted detection rate fall from 28% to 22% once the AI was taken away (Budzyń et al., 2025). These were experts. Six months of leaning on the machine eroded a skill they had spent careers building. Now picture the same dynamic where the operator is twenty-three and the AI sets every parameter.

Process knowledge is a muscle. AI that does the lifting for you does not make you stronger. It makes the muscle disappear — silently, and exactly where you will one day need it most.

Why this is a systems failure, not an HR one

Ashby's Law of Requisite Variety holds that only variety can absorb variety. On a factory floor, the experienced operator is the high-variety sensor. AI handles the routine 95% beautifully. But every time you let it deskill the human, you drain the variety reserve you keep for the other 5% — the subtle, novel, 2am fault that is the reason you employ people who can think. You optimise the average and bankrupt the exception.

The Oracle vs the Coach — a design choice, not an accident

What doesn't work: the oracle. AI as an answer-machine — it tells the operator the setting, the operator obeys, and over months the operator forgets why. Where a system turns operators into button-pushers: when the AI is wrong, nobody can tell; when it is right, nobody learns why. FOGBANK in slow motion.

What works: the coach. Junior–senior tandems where AI accelerates tacit-knowledge transfer — surfacing the senior's reasoning while the human still does the learning. And deliberately keeping the muscle: Toyota has skilled workers build components by hand, even when machines could, so the organisation never forgets the craft.

Where the Homeostat comes in

This is the design rule at the heart of Digital Kaizen: every AI intervention should be capacity-building by default. The homeostat — the self-correcting loop between people, process and technology — only stays stable if each cycle leaves people more able than the last. The one question no business case asks: does this make my people more capable, or less? Pointed the right way, the homeostat ratchets up. Pointed the wrong way, it ratchets down into elegant, efficient, brittle dependency.

Five things to check before the next system goes live

1. Ask the capacity question out loud. Does this build or erode the skill underneath it?
2. Make the AI show its reasoning, not just its answer. Seeing why it recommends 180°C is learning; seeing only "180°C" is forgetting.
3. Protect deliberate practice. Keep humans doing the hard thing manually, periodically — even when the machine could.
4. Watch for the FOGBANK gap. Map which critical process knowledge lives in one or two heads. That is your atrophy risk register.
5. Measure the exception, not just the average. Track how the floor performs when the AI is wrong — or down.