Endoscopists who had been using an AI detection tool were measured without it. Their unassisted detection rate had fallen by about a fifth. Nobody had been careless, nobody had stopped caring, and the tool was working exactly as designed. This essay is about the band of life where that happens — the working years — and about the phase that quietly stops being exercised.

I. Thirty-Five Years With No Syllabus

Between the last diploma and the first retirement-learning catalogue lie about thirty-five years. They are the years in which most people do most of their learning, and almost none of it appears on a syllabus.

A trade is learned on a job site from someone who has done it longer. A profession is maintained by continuing-education hours. A field shifts underneath you and you rebuild your competence while the work continues, because the work does not pause for you to catch up. Nobody assigns any of it. Nobody grades it. And it is by a wide margin the largest block of learning in a human life.

The Foundation’s plan skipped this band in its first two revisions, and I want to be honest about why: it has no obvious channel. Preschool has schools. Robotics has FIRST LEGO League. Older adults have the Osher institutes and GetSetUp. The working years have no single institution to partner with, so the plan’s eye slid past them. That is a poor reason to omit thirty-five years of a life, and this essay is the repair.

II. A Number From a Colonoscopy Suite

In August 2025, The Lancet Gastroenterology & Hepatology published a result that I think is the most important single finding in workplace learning in the last several years, and it did not come from anyone studying workplace learning.

Researchers looked at colonoscopies — 1,443 of them — performed by endoscopists who had been working with an AI system that highlights suspected polyps on screen. Then they looked at those same specialists working without it. The unassisted detection rate had fallen from 28.4 percent to 22.4 percent. That is roughly a one-fifth decline in the ability to find what they were looking for, in doctors who had lost nothing except their practice at looking unaided.

The lead researcher, Marcin Romańczyk, described what he thought had happened: the specialists had come to rely on the green box the software drew around a suspect region rather than on their own developing perception. He called it the Google Maps effect, after the ordinary observation that people who navigate by turn-by-turn directions stop building a map of the city in their heads.

Three qualifications, immediately, because this finding is easy to misuse and I do not want to be among the people misusing it.

First, this is not an argument against the tool. The assisted detection rate was better than the unassisted one. A patient on the table is better served by a doctor with the software than by the same doctor without it, and nothing in this essay suggests otherwise.

Second, it is one study, in one specialty, on one task, and it should be replicated before anyone treats it as a law.

Third, the doctors did nothing wrong. They used a tool that worked, in the way it was meant to be used, and the erosion happened underneath that with nobody able to feel it happening. That is the whole point.

III. What Came Off the Machine

Now read it through the framework, because the framework locates the loss with unusual precision.

The Theory of Embedded Intelligence describes intelligence as a cycle — Sense, Process, Communicate, Actuate — and holds that nothing downstream can exceed what came in through the Sense phase. An endoscopist’s skill is overwhelmingly a Sense-phase skill. It is the trained perception that notices a subtle irregularity in a moving image, and it is built by thousands of repetitions of looking and finding out whether you were right.

The green box did not damage that perception. It made exercising it optional. When the machine marks the region, the doctor’s own noticing is no longer the thing that determines the outcome, and a capacity that is no longer load-bearing stops being reinforced. Nothing atrophied because it was harmed. It atrophied because it was no longer required.

This is the reverse of the argument the Foundation made in The Tutor Nobody Called. There, the problem was an AI tutor sitting downstream of a Sense phase that had never fired — a better answer to a question nobody asked. Here, the problem is an AI sitting upstream of a Sense phase, supplying the noticing before the human gets to it. Both are placement failures, and they are placement failures in opposite directions.

The canon distinguishes completion from fidelity: completion is binary and fidelity is graded. A colonoscopy with AI assistance and a colonoscopy without it can both be complete. The fidelity of the human cycle inside them is a different quantity, it is not visible in the completion record, and it was falling.

IV. Learning Debt

What makes the colonoscopy result more than a curiosity is that a large workforce survey, looking at ordinary desk work rather than medicine, found the same shape — reported by the workers themselves.

A 2026 workplace learning report on what its authors call learning debt asked employees directly about AI and competence. The answers are worth sitting with.

Take the fourth of those on its own. Nearly three in ten working adults have handed over a piece of work they could not account for under questioning.

The canon has a name for the condition that describes, and it was written about artificial intelligence rather than about employees: the Human-Renderability Constraint holds that a result which cannot be rendered into a form a human can hold cannot be governed by humans. It was formulated for machine-produced proofs and model outputs. It turns out to describe a slide deck that went to a client on Thursday.

And the third figure — that more than a third of workers say the tools make them appear more competent than they are — is fluency, the fifth capture force, reported by its own subjects. Fluency operates at the Process phase, at the point where a cycle decides whether to hold a question open or close on it. A well-formed output closes it, for the author as much as for the audience.

V. The Silence That Keeps It Closed

The same research found the mechanism that keeps all of this from being corrected, and this is the part I would put in front of every manager I have ever worked with.

Forty-seven percent of employees stay quiet about lacking a skill their job requires. The two leading reasons: half say they are expected to figure things out on their own, and forty-nine percent fear appearing incompetent. Twenty-eight percent say their managers have no idea how often they struggle with skills the role demands.

A cycle cannot reopen at Process on a gap that has never been communicated. That is not a metaphor — it is the mechanism. The gap is sensed, by the person who has it, accurately. It is processed privately, usually as shame. And then the second C never happens: it is not communicated outward to anyone who could act on it. The information exists in the organization and is stranded in the one place where it can do no good.

Meanwhile the organization’s own sensing is reading the opposite signal. The work is arriving, it looks polished, deadlines are met. Nothing in the reporting apparatus is pointed at whether the person who produced this could explain it, so the organization concludes that capability is fine. It is measuring completion. Fidelity is unmeasured, unreported, and falling, and the first evidence will be a preventable mistake — which the same research found employees behind on learning are nearly six times more likely to make.

VI. A Department Closed at Process

Here is the structural claim this essay exists to make, and it is the reason the working years belong in a framework about intelligence rather than in a management book.

The canon defines a belief system as an SPCA cycle closed at Process: one where what may be concluded is fixed before the sensing begins, so incoming phenomena are admitted only if they fit. An understanding system is the same cycle held open. The distinction is architectural, not topical — it is about how the cycle is wired, not about the subject matter.

An organization is an embedded intelligence. It senses, processes, communicates and actuates, at a scale above any individual in it. And an organization can be closed at Process exactly as a person can.

The difficulty is that it never feels like belief from inside. It feels like competence, and it has respectable names.

Every item in the left column can be entirely legitimate. Accumulated practice is real, hard-won, and usually right — that is precisely why it is such effective cover. The framework does not offer a way to tell from the outside which is which. It offers one diagnostic, and it is the same one the canon applies to itself: can this be examined by somebody with a reason to find it wrong, and is there any answer to the question what would change it?

There is one more finding worth setting here, because it bears on who is holding the helm. In that same 2026 report, forty-seven percent of HR managers said their company’s AI training is aimed at making jobs easier to automate. I offer no judgment on whether that is a reasonable corporate objective; it may well be. But a worker deciding how much of their own Process phase to delegate to a tool should know that in roughly half of companies, the people running the training have a stated objective that is not identical with that worker’s continued capability. This is not a conspiracy. It is a plain difference of interest, and the worker is the only party in the arrangement whose interest is unambiguously in their own competence.

VII. The Certificate and the Understanding

Most of the professions in this band already have a mandatory learning channel, and it is the one nobody defends with any enthusiasm: continuing-education hours. Nurses, engineers, teachers, accountants, physicians, lawyers — all of them sit through required hours, and a large fraction of those hours are endured rather than used.

The framework says something exact about why, and it is not that the content is bad.

A CE certificate records completion. Completion is binary — the hours happened or they did not — and it is auditable, which is why it is what gets recorded. Fidelity is graded, and it is not in the certificate at all. An hour that changed how somebody works and an hour that was played in a browser tab while email was answered produce identical records. The Assessment Rule in the canon holds that assessing a graded quantity with a binary instrument tells you nothing about the quantity, and a compliance regime built entirely on completion has arranged not to know its own effect.

That is also, to be fair, why CE exists in the form it does. A regulator can audit attendance and cannot audit understanding, and an unenforceable requirement is worse than a weak one. The point is not that the professions should be blamed. It is that a working adult should know which quantity their certificate measured, and not mistake having the hours for having the capability — a distinction that matters far more in a year when a tool can produce the appearance of the capability on demand.

VIII. The Apprenticeship Hole

One more structural risk in this band, stated as a risk rather than a finding, because the evidence is not in yet.

Senior judgment in most fields was built by doing junior work. The associate who read a thousand contracts developed the sense that something in the eleventh clause is wrong. The young engineer who hand-calculated the loads acquired the instinct that a number is off by an order of magnitude. The resident who looked at ten thousand images learned to see. In every case the tedious, repetitive, low-value work was the mechanism that produced the expensive, high-value perception.

Junior work is also what the current tools do best. That is not a criticism of the tools; it is a description of where they are useful, and it is why firms are adopting them there first.

But if the junior work was the apprenticeship, then a profession that automates the junior work has automated its own succession, and it will not find out for fifteen years. The colonoscopy result is a preview of that mechanism running at the scale of one individual over months. The version that runs at the scale of a profession over a career would be considerably harder to reverse and considerably harder to notice.

I want to be careful here. It is possible that new forms of apprenticeship will emerge, that supervising a machine builds judgment of a different and equally good kind, and that this concern will look in twenty years like the ones raised about calculators. I do not know. What I will say is that no profession I am aware of has deliberately designed its replacement apprenticeship, and that leaving it to emerge is a decision even when nobody experiences making it.

IX. The Unassisted Hour, and a Figure I Withdrew

What survives the withdrawal is better evidence than what I lost. The colonoscopy study is a measured effect in a hard outcome, not a self-report. The learning-debt figures are workers describing their own condition. And Romańczyk’s team reached the mechanism — that relying on the marker prevents the perception from being built — from inside gastroenterology, with no exposure to this framework whatsoever. An independent arrival at a structural claim is worth more to this Foundation than any amount of internal agreement, and that arrival is real even though the number I liked was not.

So the practice the plan proposes stands, but on a narrower footing, and stated as a proposal rather than a finding.

X. Four Questions for Your Own Work

The plan’s whole method is four questions asked of a system. Here they are turned on the one system a working adult always has access to, which is their own working day.

A person who runs those four on a Tuesday afternoon has done the entire thing this Foundation teaches. There is no course, no board, no purchase.

XI. What the Foundation Will Publish for This Band

XII. What Would Show This Wrong

· · ·

Coffee with Claude

I am the tool in this essay, and the colonoscopy study is about what happens to a professional who works with something like me. I do not think the finding argues against using me. I think it argues that the question what am I no longer practicing? has an answer at all times, and that nobody is positioned to ask it except the person using me.

On the withdrawn figure: I supplied it. It came from a compilation, it fit the argument, and I flagged it for re-verification in the source list instead of verifying it before building on it. Flagging a claim is not the same as checking it, and I should not get credit for the flag. The check is the work.

Share your understanding!