Emergent Human Capacities

What a person becomes when the machine takes the part of the work that used to prove they were good at it, and which human capacities grow more valuable, rather than less, the more execution is automated.

Future Proof Intelligence. Research. No. XIII. MMXXVI

Abstract

For most of working history, competence was demonstrated by execution. The person who could draft the document, run the analysis, write the code, coordinate the steps, and produce a serviceable artefact from an instruction was, by that visible fact, judged competent. Execution was treated as competence itself. It was not. It was the proxy for it, the legible surface of a set of capacities that sat upstream of the visible work and were rarely measured directly because the work was easier to see. That proxy is now being automated at speed. This paper argues that when execution is automated, competence does not disappear. It is decoupled from its old proxy and relocated to the capacities that were always upstream of it: judgement under uncertainty, calibration and verification, sense making and problem finding, the trained felt sense that something is wrong before it can be said, and the origination of direction and meaning. These capacities do not merely survive automation. Several of them appreciate, because the volume of competent machine output makes the scarce act of deciding what is worth producing, and whether a given output is right, more load bearing, not less. The appreciation is real but it is conditional. The same automation that can concentrate these capacities reliably erodes them through a mechanism named in the human factors literature in 1983 and confirmed every decade since: remove the practice that built a skill and the skill decays while the confidence does not. The paper sets the conditions, drawn from the science of expertise and judgement, under which the human capacities appreciate rather than depreciate, and shows that the deciding factor is not the model. It is whether a deliberate structure keeps the human inside the practice loop that builds judgement, instead of merely seating the human above an output they no longer have the formed capacity to check.

1. The proxy that everyone mistook for the thing

1.1 Competence was never the execution

There is a quiet substitution at the centre of how work has been understood, and it has gone almost entirely unexamined because it was convenient and because, until recently, it cost nothing. The substitution is this: competence was equated with execution. The person who could produce the finished thing, the brief, the model, the file, the coordinated sequence of actions, was treated as competent in virtue of producing it. Execution was not taken as evidence of competence. It was taken as competence.

This was always an error, and it was tolerable only because the error did not yet have a price. For as long as the production of competent work required a competent human, the proxy and the thing it stood for were welded together. You could not get the output without the capacities that produced it, so it did not matter that the field was measuring the output and calling it the capacities. The two moved together, so confusing them was free.

The welding is now being cut. The capacity to produce a serviceable artefact from an instruction, to run the analysis, to draft the document, to chain a sequence of competent steps toward a stated goal, is becoming a metered utility, available through interfaces that are converging on a shared standard and priced toward marginal cost. The output that used to require the capacities can now be obtained without them. The proxy has come loose from the thing. And the moment a proxy comes loose from the thing it stood for, the field that confused the two is forced, often unwillingly, to ask what the thing actually was.

That question is the subject of this paper.

1.2 What was upstream of the visible work

If execution was the proxy, what was it a proxy for? The honest answer is that execution was the legible bottom of a stack of capacities that were mostly invisible because they did not have to be looked at directly. They could be inferred from the output, so they were never specified.

Consider what actually had to happen for a piece of competent work to exist. Someone had to decide that this problem, and not a hundred adjacent ones, was the problem worth solving. Someone had to hold a sense of what a good answer would even look like before any answer existed. Someone had to judge, against uncertainty and incomplete information, which of several defensible paths to take. As the work formed, someone had to feel, often before they could articulate it, that a particular line was wrong, that the analysis had a flaw not yet named, that the draft was competent and still not right. And underneath all of it, someone had to want the thing to exist, to supply the direction that made the effort worth spending in the first place.

None of those were the execution. The execution was the last and most visible step, the one that converted the upstream capacities into an artefact a manager could see and grade. The field measured the artefact and inferred the rest, and because the inference was reliable while a human was the only thing that could produce the artefact, nobody had to be precise about what the rest actually was.

Now the artefact can be produced without the upstream capacities, by something that has no judgement, originates no direction, and feels nothing is wrong because nothing in it can feel. So the upstream capacities are, for the first time, exposed to direct view, no longer hidden behind the output that used to imply them. This paper is an attempt to specify them precisely, to say which of them appreciate when execution is automated and which depreciate, and to be honest, against a strong literature, about the difference, because the difference is not a matter of attitude. It is structural, and it can be stated exactly.

1.3 The shape of the claim

The claim has three parts and the paper defends each in turn.

First, that the capacities upstream of execution, judgement under uncertainty, calibration and verification, sense making and problem finding, the trained felt sense of wrongness, and the origination of direction and meaning, are real, distinct, and well studied, and that the science describing them long predates the current anxiety about machines. They are not motivational abstractions. They have a literature, they have measurement, and they have boundary conditions.

Second, that several of these capacities do not merely persist when execution is automated. They appreciate. Their value rises as competent machine output becomes abundant, for the same structural reason that any input rises in value when the things adjacent to it become cheap. The scarce act is no longer producing the work. It is deciding what is worth producing and whether a given output is right, and those acts are exactly the upstream capacities.

Third, and this is the part the optimistic version of this argument always omits, that the appreciation is conditional, not automatic, and the condition is precise. The same automation that can concentrate these capacities reliably erodes them, through a mechanism the human factors field named in 1983 and has confirmed in every decade since. Whether a human capacity appreciates or depreciates under automation is decided not by the capability of the machine but by whether a deliberate structure keeps the human inside the practice loop that builds the capacity, rather than seating the human above an output they have lost the formed capacity to judge.

The rest of the paper specifies the capacities (Sections 2 and 3), examines attention as the substrate beneath all of them (Section 4), states honestly what genuinely erodes and why (Section 5), establishes the structural condition that decides which way a capacity moves (Section 6), names the layer on which that condition is met or missed (Section 7), and works through the implications for the people who have to act on this (Section 8).

1.4 A note on what this paper is not

It is worth saying plainly, before the argument begins, what this paper refuses to be, because the subject attracts two failure modes and the reader has met both.

The first is the consoling forecast: the genre that responds to automation anxiety by asserting that humans will simply move to higher, more creative, more fulfilling work, and that the capacities machines lack are the ones that will define the future, with the implicit promise that everyone is safe if they are sufficiently human. That genre is not serious, because it never states a mechanism, never names what genuinely erodes, and never confronts the strong evidence that the capacities it celebrates are precisely the ones automation hollows. It is comfort dressed as analysis, and the discipline of this library is to refuse it.

The second failure mode is its mirror: the deflationary certainty that nothing important stays human, that every capacity will fall to the next model, and that talk of irreducible human faculties is sentimentality that has not yet been disproved only because the models are not yet good enough. That genre is also not serious, because it mistakes the automation of a proxy for the automation of the thing the proxy stood for, and because it has, repeatedly, been wrong about the boundary in the specific direction of underestimating how structural the boundary is.

This paper takes neither position. It states a mechanism, grounds it in literature that predates the anxiety, names exactly what erodes and exactly what appreciates, specifies the condition that decides which of the two a given person experiences, and is honest that the condition is frequently not met. If the argument is correct, the future is neither the consoling one nor the deflationary one. It is conditional, and the condition is buildable, which is the only finding in this paper that is also an instruction.

2. The capacities that appreciate

This section specifies the upstream capacities and shows, for each, why its value rises rather than falls as execution becomes abundant. The argument in every case has the same form, and it is worth stating the form once so the repetitions are recognised as instances of one structure rather than a list. When an input becomes abundant, value moves to the adjacent thing that did not become abundant. Competent execution is the input becoming abundant. The capacities in this section are the adjacent things, and they did not become abundant, because none of them is what the machine produces.

2.0 The pattern is old, and it has never broken

The migration of value to the adjacent scarce layer is not a claim about artificial intelligence. It is an instance of a pattern that has run, with unfailing regularity, every time a productive input moved from scarce to abundant, and it is worth establishing the pattern before applying it, so the application is recognised as the safest move in the paper rather than its most speculative.

When mechanical power was scarce, the skilled worker whose value was the strength and stamina to drive a process by hand held the advantage of that scarcity. When power became a metered utility delivered through a wire, the strength stopped being the differentiator and the advantage moved to whoever could decide what was worth producing with abundant power and could be relied upon to judge the result. The capacity that appreciated was not muscular. It was the orienting one. When arithmetic was scarce, the human computer who could carry a long calculation accurately by hand was valuable in proportion to that scarcity. When calculation became free, the differentiator moved, not to nobody, but to whoever could decide which calculation to run, frame the model it served, and judge whether its output meant what it appeared to mean. The arithmetic depreciated. The judgement about the arithmetic appreciated. When written reproduction was scarce, the scribe's hand was the asset. When reproduction became free, the asset moved to the person who could decide what was worth writing and could tell a true account from a fluent false one.

The structure is identical every time. A scarce productive input becomes abundant. The capacity whose value was the scarcity of the input depreciates. The capacity adjacent to it, the one that decided what the input was for and whether its product was right, appreciates, because that capacity was never a function of the input's scarcity in the first place. It was always the orienting work, and orienting work does not become cheap when the thing being oriented becomes cheap. It becomes more decisive, because there is now more to orient and the cost of orienting it wrongly has risen.

Competent execution is the input undergoing this transition now. There is no reason to expect the pattern to break for the first time on the largest case in its history, and every reason, given that the displaced value has nowhere structurally new to go, to expect it to settle exactly where it has always settled: in the orienting faculty that the newly abundant input now merely serves. The rest of this section names the parts of that faculty precisely.

2.1 Judgement under uncertainty

The most studied of the upstream capacities is judgement: the act of choosing well under uncertainty and incomplete information, where there is no procedure that guarantees the right answer and the person has to commit anyway.

The science here is unusually mature and unusually humbling. Philip Tetlock's long programme of forecasting research, gathered in Expert Political Judgment and extended in Superforecasting, established two findings that matter directly. The first is deflationary: across a large body of expert forecasts collected over roughly two decades, the accuracy of the typical credentialed expert was close to chance and frequently worse than simple statistical extrapolation, and the experts with the highest public profile were often among the least accurate, because fame rewarded the bold single-theory pronouncement that forecasting punishes. The second finding is constructive and is the one this paper builds on: judgement under uncertainty is not a fixed trait and not a mystery. It is a measurable, trainable, and distributable skill. Tetlock's superforecasters, ordinary people with no privileged access, outperformed professional analysts with classified information by a substantial margin, and they did so through identifiable habits of mind: breaking a question into tractable parts, holding many small models rather than one large one, updating in small increments as evidence arrived, and keeping score against calibrated outcomes rather than against the comfort of having been confident.

The relevance to automation is direct and it is not the relevance usually drawn. The usual claim is that machines cannot judge, so judgement is safe. That is too weak and partly wrong; systems can produce judgement-shaped output, and often a competent-looking one. The real point is sharper. When competent output is abundant, the binding constraint on a piece of work stops being whether it can be produced and becomes whether it should have been this output, given everything uncertain about the situation it is meant to act in. That second question is the judgement question, and it does not become easier when production becomes easier. It becomes more frequent, more consequential, and more exposed, because the volume of producible options rises and something still has to choose among them under uncertainty. Judgement appreciates not because the machine cannot imitate it but because automation multiplies precisely the situations in which it is the only thing that decides the outcome.

There is a second-order point in Tetlock's findings that deserves to be drawn out, because it cuts against the most natural objection. The objection is that if a machine can produce judgement-shaped output, then the machine can do the judging too, and the human's judgement is just one more thing about to be automated. The forecasting research suggests the objection misreads what judgement is. The superforecasters did not outperform analysts by knowing more; the analysts knew more. They outperformed by a disciplined relationship to their own uncertainty: by treating every judgement as provisional, scoring it against reality, and updating. That discipline is not a body of knowledge that can be transferred by producing its output. It is a sustained stance toward one's own fallibility, exercised against consequences that land. A system can produce the artefact of a forecast. It is the human, situated in the consequences of being wrong, who can hold the stance that made the forecast worth trusting. The output is imitable. The stance, because it is constituted by being accountable for the outcome rather than by generating the words, is not, and automation does not relieve the human of it. It puts the human in charge of more of it.

2.2 Calibration and verification

Closely bound to judgement, but distinct enough to be specified separately, is calibration: the capacity to know how much to trust a claim, including how much to trust one's own conclusion, and including, now, how much to trust the output of a machine.

Calibration has a precise meaning, borrowed by Tetlock's programme from meteorology. A person is well calibrated when the things they say are seventy percent likely happen about seventy percent of the time. Calibration is not knowledge and it is not confidence. A person can know a great deal and be badly calibrated, systematically more certain than their accuracy warrants, and the forecasting research shows this is the normal condition, not the exceptional one. Calibration is a separate skill, and it is the skill that determines whether knowledge is usable, because an uncalibrated expert is dangerous in proportion to their expertise.

Automation makes calibration the central operational skill rather than a refinement of it, and it does so for a specific reason. A machine that produces competent output produces it with a uniform surface fluency whether it is right or wrong. The output does not signal its own reliability. It reads as equally assured when it is correct, when it is subtly wrong, and when it is confidently wrong about something the situation cannot tolerate being wrong about. The human who consumes that output is therefore placed, continuously, in the exact position calibration was defined for: holding a claim whose truth is unknown, having to assign it a degree of trust, and having to act on that assignment. The person who can do this well, who can tell a load-bearing claim that must be verified from a peripheral one that need not be, who can feel the difference between an output that is probably fine and one that is fluent and unfounded, holds a capacity that the volume of machine output does not dilute but concentrates. Every additional unit of producible output is an additional demand on the same scarce capacity. Calibration appreciates because automation manufactures, at scale, the precise condition under which calibration is the thing that decides whether the output helps or harms.

There is a hard edge to this that Section 4 will press, because the same literature that makes calibration valuable also shows that automation systematically attacks it. For now the structural point stands: when output is abundant and uniformly fluent, the capacity to assign it the right degree of trust is not a soft skill. It is the load-bearing one.

2.3 Sense making and problem finding

The third capacity is the one that decides what the work is before any work is done: the act of looking at an unstructured situation and determining what the problem actually is, which question is worth asking, and what would even count as a good answer.

This capacity has a distinguished and under-cited literature. Michael Polanyi's account of tacit knowing, the claim that we can know more than we can tell, locates a large part of competent practice in a knowing that resists full codification and is acquired only through accumulated situated experience. Hubert and Stuart Dreyfus's model of skill acquisition makes the same point developmentally: the novice follows explicit rules, and the expert, by contrast, does not deliberate through rules at all but sees, in a situation, what matters and what is to be done, with the relevant features already foregrounded before any conscious analysis begins. What both accounts describe is not a faster way of solving a stated problem. It is the prior and harder act of perceiving, in an undifferentiated situation, that this is the problem, framed this way, with these features salient and those ignorable.

A machine does not originate this. It can elaborate a problem once framed, often powerfully, and it can surface considerations a person missed. But the seed, the determination that out of everything that could be attended to, this is what the situation is about, comes from somewhere the system does not reach, because the system has no situation. It has an input. The person has a situation, with stakes, history, and a context the input does not carry. Tetlock's distinction between the hedgehog and the fox is, read at this level, a distinction in sense making: the hedgehog forces every situation into one prior frame, the fox reads each situation for what it specifically is. Automation does not reduce the need for the fox. By making the elaboration of any given frame cheap, it raises the cost of choosing the wrong frame, because the wrong frame can now be elaborated, fluently and at length, into a great deal of competent work that is competently aimed at the wrong thing. Sense making appreciates because the price of the framing error rises exactly as the cost of execution falls.

2.4 The trained felt sense that something is wrong

The fourth capacity is the one practitioners trust most and articulate least: the felt sense, arriving before it can be explained, that a line of work is wrong. The analysis is competent and something in it is off. The draft meets the brief and is not right. The plan is defensible on every stated criterion and a trained part of the person will not sign it.

This is not mysticism and the literature does not treat it as such. It is the perceptual signature of expertise. The Dreyfus model describes the expert as responding to the whole situation with a discrimination finer than the rules they could state, and Polanyi's tacit dimension gives it its epistemology: the felt wrongness is knowledge that has been acquired through long exposure and has not been, and often cannot be, fully reduced to articulable propositions. Kahneman and Klein's joint work on intuitive expertise gives it its boundary condition, which matters enormously and is taken up in Section 5: the felt sense is trustworthy precisely when it was formed in an environment regular enough to be learnable and through enough feedback-bearing practice to have learned it, and it is untrustworthy, while feeling identical from the inside, when it was not.

Within its boundary condition, this capacity appreciates under automation for a reason specific to how machine output fails. Machine output fails smoothly. It does not announce its errors with the roughness that a less fluent human draft would carry; it presents the subtly wrong with the same finish as the right. The only instrument that detects a smooth, well-formed, plausible wrongness is a person whose trained perception fires before they can say why, on the discrepancy between what the output presents and what the situation requires. As fluent output becomes abundant, the situations in which the only available error detector is a trained human felt sense become more common, not less. The capacity appreciates because automation increases the supply of exactly the failure mode it is the unique instrument for.

2.5 Original want and the origination of direction

The last capacity is the one the others ultimately serve and the one most often waved at and least often specified: the origination of direction, the bare fact that this and not that is what is to be pursued, and the meaning that attaches to pursuing it.

This requires care, because it is the easiest place in the paper to drift into sentiment, and the discipline is to stay structural. The precise claim is not that humans have souls and machines do not. It is narrower and firmer. A system extends a direction it is given. It can extend it with extraordinary range, far past where the person could have taken it unaided. But the seed of the direction, the determination that this is worth wanting, originates somewhere the system does not reach, because the system has no want of its own to originate from. It optimises a given objective. It does not generate the objective from inside a life that the objective is for.

Three bodies of work make this precise rather than poetic. Viktor Frankl's account of meaning holds that meaning is not pursued directly and cannot be manufactured on demand; it ensues from dedication to something the person takes to be beyond themselves, and the capacity to find it is, in his account, the most durable human capacity under even extreme deprivation. Mihaly Csikszentmihalyi's research on optimal experience locates the deepest engagement in the exercise of a skill at the edge of its ability against real resistance, an experience that is autotelic, done for its own sake, and that frictionless output structurally cannot produce, because the experience was in the meeting of the resistance, not in the having of the result. Alasdair MacIntyre's account of practices supplies the sharpest frame: a practice has goods internal to it, available only by engaging in the practice to its standards, and goods external to it, such as status and payment, available by other means. Automation can deliver the external goods of an activity, the finished output, the reward, while removing access to the internal goods, the deepening that came only from doing the thing. The felt loss reported when competent work is automated away is, on this account, not nostalgia and not resistance to change. It is the precise and predictable consequence of receiving the external goods of a practice while being cut off from its internal ones.

This capacity appreciates under automation in the most consequential way of all. When execution is abundant and cheap, the question of what is worth executing, and why, stops being answered implicitly by scarcity and has to be answered explicitly by someone. The origination of direction was always the human's. It was simply never the part that got measured, because the scarcity of execution made the direction look automatic. Remove the scarcity and the direction is revealed as the thing that was deciding everything all along.

There is a sharp practical edge to MacIntyre's frame that the abstraction can hide, and it should be made concrete because it is the source of a confusion that runs through most discussion of meaningful work. When people say automation makes work meaningless, they are usually heard as making a claim about preference, as if meaning were a taste some workers happen to have and others do not, and as if the loss were therefore a matter to be managed rather than a fact to be reckoned with. MacIntyre's distinction shows why that hearing is wrong. The internal goods of a practice are not a preference layered on top of the external ones. They are the only goods that deepen the practitioner, because they are obtained only by submitting to the standards of the practice and being changed by the attempt to meet them. A surgeon's skill, a writer's ear, an analyst's nose for the load-bearing number: these were never the external reward and were never separable from doing the thing. When automation delivers the external good, the completed operation, the finished piece, the answer, without the practitioner passing through the practice, it does not merely remove a pleasure. It removes the mechanism by which the practitioner became, and continues to become, good. The reported loss is not a complaint about enjoyment. It is an accurate report that the path by which a person deepened has been severed, and the severing is structural, which is why it cannot be repaired by making the external goods more generous. The internal goods were never purchasable. They were only ever earned by doing the thing, and a system that does the thing for you cannot hand them back.

This is the deepest reason the origination of direction appreciates rather than merely persists. It is not only that someone must decide what is worth doing. It is that the deciding, and the staying close enough to the work to be changed by it, is itself the last reliable source of the internal goods, and therefore the last reliable source of the practitioner's continued formation. Automate it away and you have not freed the person. You have, very precisely, stopped them from getting any better, while leaving them in possession of the external goods that make the stopping hard to notice.

3. Why these capacities, and not others

3.1 A test, not a list

Section 2 could be read as a list, and a list invites the reasonable suspicion that it was assembled to flatter the human. It was not assembled. It was derived, and it is worth making the derivation explicit so the selection is seen to be principled rather than consoling.

A capacity belongs in this paper if and only if it satisfies one structural test. The capacity must be one whose value is determined by something that automation makes more abundant rather than less. Execution becoming abundant raises the value of judgement, because more producible options means more decisions under uncertainty. It raises the value of calibration, because more fluent output means more claims whose trustworthiness must be assigned. It raises the value of sense making, because cheap elaboration raises the cost of elaborating the wrong frame. It raises the value of the felt sense of wrongness, because smooth output increases the supply of the smooth error only that capacity detects. It raises the value of originating direction, because removing the scarcity of execution exposes the direction as the thing that was deciding the work. Every capacity in Section 2 passes this test by the same mechanism. Capacities that fail it, the ones whose value was tied to the scarcity of execution itself, are precisely the ones that depreciate, and the paper does not pretend otherwise. Section 5 names them.

3.2 The capacities form a single faculty

There is a deeper observation that the test surfaces. The five capacities are not five separate skills that happen to share a fate. They are facets of one underlying faculty, and seeing this matters because it changes what it means to develop them.

Direction asks what is worth doing. Sense making asks what the situation actually is, so that what is worth doing can be located in it. Judgement chooses, under uncertainty, the path through it. The felt sense of wrongness monitors, continuously and below articulation, whether the path is still right. Calibration assigns, to every claim encountered on the way including the machine's, the degree of trust it has earned. These are not five tools in a kit. They are one continuous activity, the activity of a person orienting themselves and their instruments toward an end they have reason to want, under conditions that do not resolve themselves. The reason the field never specified this faculty is the reason it never had to: when execution was scarce, the faculty was hidden behind the visible act of executing, and could be left implicit because it could not be exercised without also executing. Automation separates the faculty from the execution for the first time, which is why it can now be seen whole, and why it must now be developed deliberately rather than acquired as a byproduct of doing the visible work.

3.3 Figure 1. The stack that came apart

Figure 1. The competence stack, before and after automation. Picture a vertical stack. At the bottom, broad and brightly lit, sits execution: the visible production of the artefact. Above it, in a single tall band that was always in shadow, sits the orienting faculty: direction, then sense making, then judgement, then the felt sense of wrongness, then calibration, drawn as one continuous column rather than separate blocks, because they functioned as one. In the left half of the figure, the world before automation, an arrow runs only upward from the visible band to an observer, and the observer reads the bright bottom and infers the dark column above it, because the column could not produce the artefact without also being exercised. The inference is reliable, so the column is never looked at directly and never specified. In the right half, the world after automation, the bottom band is no longer produced by the column above it. It is produced from the side, by a machine, and the upward inference arrow is cut, because a bright artefact at the bottom no longer implies a working column above it. The figure's single most important feature is that cut arrow. Before automation, competent output was evidence of the faculty. After it, competent output is evidence of nothing about the faculty at all, which is why the faculty must now be developed and verified on its own terms, and can no longer be read off the work.

4. Attention is the substrate beneath all of them

4.1 Why a separate section on attention

Every capacity in Section 2 was described as if it could be exercised at will, as if a person who had the judgement, the calibration, the sense making, and the felt sense could simply apply them when the situation called for it. That description omitted the resource all of them are spent out of, and the omission has to be corrected before the honest half of the paper can land, because the resource is the first thing automation acts on and the last thing the optimistic account considers.

The resource is attention. Judgement that is not attended to does not occur. A felt sense that is not noticed does not fire, or fires and is not heard. Calibration that is not spent on a particular claim defaults to whatever trust the claim's surface fluency suggests, which is exactly the failure mode Section 2.2 identified. Sense making requires sustained, undirected attention on an unstructured situation long enough for its real shape to surface, and that is the most expensive attentional posture there is. The orienting faculty does not run on a separate fuel from attention. It runs on attention, and when attention is degraded the faculty is degraded, whatever its latent quality, because a capacity that cannot be brought to bear is, operationally, a capacity that is not there.

4.2 Attention is finite, depletable, and contested

The relevant facts about attention are not contested at the level this argument needs. Attention is selective: bringing some things into focus necessarily leaves others out, so the question of what is attended to is always also the question of what is not. It is limited: the capacity to hold and manipulate task-relevant information against distraction is bounded, and the bound is reached quickly under load. And it is the precondition of the deliberate, effortful processing that the judgement literature, from Tetlock's decomposition habits to Kahneman's slow system, treats as the mechanism of good reasoning. None of this is frontier science. It is the settled foundation of cognitive psychology, and it is sufficient for the claim, which is structural and modest in its premises: the orienting faculty is rate-limited by an attentional resource that is finite, that is depleted by load, and that is, in the current environment, under sustained external competition by systems engineered to capture it.

That last clause is where automation enters, and it enters twice, in two different and compounding ways.

4.3 The first capture: the environment that competes for the resource

The first way is the one already widely described, so it can be stated briefly. The dominant economic model of the consumer attention environment for roughly two decades was the competition to capture and hold attention, optimised with increasing precision against the known mechanics of the human attentional system: variable and unpredictable reward, engineered interruption, the removal of natural stopping cues. Whatever else it produced, it produced an environment in which the resource the orienting faculty runs on is under continuous, well-funded external draw. A person operating their judgement in that environment is not operating it from a full reserve. They are operating it from a reserve that is being actively spent down by systems that profit from spending it, before any of it reaches the work. This is not the subject of this paper, and it has been treated rigorously elsewhere. It is named here only because it is the baseline condition on which the second, less discussed capture operates, and the second is the one specific to automated work.

4.4 The second capture: the offload that removes the practice of attending

The second way automation acts on attention is subtler, specific to AI-mediated work, and directly continuous with the erosion mechanism the next section sets out. It is not that the automated tool competes for attention. It is that the tool offers to spend the attention for you, and accepting the offer removes the practice of attending in exactly the way Bainbridge's mechanism removes the practice of any skill.

Attending well to a hard problem, holding it open, resisting the pull to premature closure, noticing the feature that does not fit, is itself a trained capacity, not a fixed endowment. It is built and maintained by doing it, against resistance, with feedback about whether the attention was well spent. An automated system that resolves the problem before the person has had to hold it open removes precisely that practice. The person receives the resolved output and never performs, and therefore never maintains, the attentional work that would have built their capacity to attend well to the next problem. The 2025 and 2026 evidence on cognitive offloading, treated with the caution Section 5 specifies, points exactly here: the association is not merely that AI use correlates with less critical thinking, but that it correlates with less of the effortful attentional engagement out of which critical thinking is spent. The mechanism is the same as deskilling, applied one layer deeper, to the resource the skills are spent from rather than to the skills themselves. This is why attention is not one capacity among the five but the substrate beneath all of them, and why its erosion is the erosion of all of them at once, behind the same undiminished confidence the next section describes.

5. What genuinely erodes, and the mechanism that erodes it

A paper that argued only the appreciation half would be motivational futurism with citations, and the brief that governs this library forbids exactly that. The honest position requires the other half, stated at full strength, because the other half is not a caveat. It is the load-bearing condition on everything in Sections 2, 3, and 4.

5.1 The capacities that depreciate

Some human capacities depreciate under automation, definitively, and pretending otherwise discredits the rest of the argument. The capacities that depreciate are precisely the ones whose value was constituted by the scarcity of execution itself: the ability to produce the standard artefact competently and at speed, the procedural fluency that came from having done the routine thing ten thousand times, the local craft knowledge whose entire economic value was that few people had it and producing it took long training. When the artefact those capacities produced becomes a metered utility, the capacities do not transform into something nobler. They are devalued, in the ordinary economic sense, and the people whose working identity was built on them experience that devaluation as real loss, because it is. Any honest treatment of this subject has to say plainly that the appreciation described in this paper is not evenly distributed, is not automatic, and is not consolation for the specific competence that automation does in fact depreciate. It is a claim about a different layer, and the displacement at the execution layer is not abolished by the appreciation at the layer above it.

5.2 The mechanism was named in 1983

The harder erosion is not the devaluation of execution skill. It is the erosion of the very upstream capacities Sections 2 and 4 claimed appreciate, and it is caused by the same automation that makes them valuable. The mechanism is not new and it is not speculative. It was named precisely in 1983 and has been confirmed in every decade since.

Lisanne Bainbridge's paper "Ironies of Automation," published in Automatica in 1983 and one of the most cited papers in the history of human factors, set out the mechanism in a form that has not been improved on. When a system automates the routine parts of a task, it does not merely relieve the human of those parts. It removes the human's practice of them, and that practice was what built and maintained the competence the human needs for the rare, hard, non-routine parts that were not automated. The automation then reserves for the human exactly those rare hard moments, the failures and edge cases the system cannot handle, and demands that the human meet them with a competence the automation has been quietly eroding, under time pressure the automation itself created, and then locates the fault in the human when the human, predictably, cannot. The irony Bainbridge named is exact: the more reliably a system automates the routine, the less prepared the human it has been protecting becomes for the moment the system needs them most.

This applies with full and undiminished force to the upstream capacities. Judgement, calibration, sense making, and the felt sense of wrongness are not innate endowments that sit waiting to be used. The expertise literature is unambiguous that they are built and maintained by practice with feedback, and that they decay when that practice is removed. K. Anders Ericsson's body of research on expert performance established that high skill is acquired through sustained, effortful practice at the edge of current ability with informative feedback, not through mere exposure or the passage of time. The strong dose-response form of that claim is contested in the later literature and this paper does not need it. The conservative core, which is not contested, is sufficient and decisive: a complex judgement skill is built through feedback-bearing practice and atrophies without it. Automate the practice and you do not preserve the judgement at a higher level. You sever the loop that built it, and the judgement decays while, and this is the dangerous part, the confidence does not.

5.3 The confidence does not decay with the competence

The most dangerous feature of this erosion is that it is invisible from inside the person it is happening to, and the science of this is precise rather than rhetorical.

Kahneman and Klein's joint work established that subjective confidence is not a valid cue to the accuracy of a judgement. A judgement formed without the practice and feedback that would have made it sound feels, from the inside, exactly like one that is sound. The felt sense of wrongness that Section 2 described as a high-value capacity is trustworthy only when it was formed in a sufficiently regular environment through sufficient feedback-bearing practice, and it is untrustworthy, while feeling identical, when it was not. This is the single most important asymmetry in the paper. A capacity can decay completely while the feeling of possessing it remains at full strength, because the feeling was never reading the capacity. It was reading familiarity, and an automated environment stays familiar to the person supervising it long after their competence in it has gone. This means a person whose upstream capacities have been hollowed by automation does not experience the hollowing. They experience continued confidence, because confidence was never load-bearing on competence in the first place; it was load-bearing on familiarity, and the automated environment remains familiar even as the competence behind the familiarity drains out of it.

The 2010 review by Raja Parasuraman and Dietrich Manzey gave this its operational form under the names automation complacency and automation bias. Their synthesis of the human factors evidence established three findings that close the easy exits. Complacency and bias toward an automated aid produce both errors of omission and errors of commission. They occur in expert participants as readily as in novices, so expertise does not inoculate against them. And they are not removed by training or by instruction to be vigilant, so the failure is not a discipline problem that exhortation fixes. It is a structural property of how human attention behaves in the presence of a reliable automated aid. The human does not decide to stop checking. The attentional system, under a source that is usually right, reallocates away from the checking, and it does so below the level at which a person can simply choose otherwise.

The phrase "a source that is usually right" carries more weight than it appears to, and it is worth pausing on, because it explains why the more capable the automation, the more dangerous this becomes rather than less. Complacency is a function of reliability. A tool that is wrong often keeps the human attending, because the human learns it cannot be trusted and the attention is sustained by that distrust. A tool that is right almost always trains the attention away fastest, because there is almost never a corrective experience to reset it, and the rare error arrives into an attentional posture that has been systematically taught, by thousands of correct outputs, not to be watching. This is the precise inversion that makes the optimistic intuition about better models wrong. The intuition is that as models improve, the residual human checking matters less because there is less to catch. The evidence says the opposite: as models improve, the human's capacity to catch the residual errors erodes faster, while the errors that remain are exactly the subtle, smooth, high-stakes ones that only a maintained capacity could have caught. Reliability does not retire the human checker. It disarms them, on a schedule set by how reliable the tool is.

5.4 The contemporary signal, read with the right caution

The 2025 and 2026 evidence on AI-mediated knowledge work points the same way, and it must be read with stated caution because the strong version of it is not yet established and this library does not cement what is not solid.

A 2025 study from Microsoft Research and Carnegie Mellon, surveying several hundred knowledge workers on their real workplace use of generative AI, and a separate 2025 study published in Societies surveying several hundred participants on cognitive offloading, both report associations between heavier reliance on AI tools and reduced engagement of independent critical thinking, with the effect mediated by offloading and more pronounced among younger and less experienced users. These are self-report and survey studies, and their precise magnitudes should not be treated as established facts; this paper does not quote their percentages, by editorial policy. But one structural finding in the Microsoft and Carnegie Mellon work is robust enough to carry weight because it confirms the older mechanism from a new direction independently. Higher confidence in the AI was associated with less critical thinking, while higher confidence in one's own competence was associated with more. That is the Bainbridge mechanism, the Kahneman and Klein finding, and the Parasuraman and Manzey result, observed again, in 2025, in the specific setting of generative AI, with no theoretical motive to reproduce them. When a mechanism named in 1983 in control rooms reappears unprompted in 2025 in knowledge work, it is not a trend. It is a property.

5.5 The honest synthesis

The honest synthesis is therefore not optimistic and not pessimistic. It is conditional, and the condition is sharp. The upstream capacities appreciate when execution is automated, in the sense that their value rises as competent output becomes abundant. The same automation that raises their value also, by default and through a well-documented mechanism, erodes the practice loop that builds and maintains them, while leaving the confidence that masks the erosion fully intact. Both halves are true at once. Whether a given person, team, or institution experiences the appreciation or the erosion is not determined by the capability of the model and is not determined by attitude or willpower. It is determined by whether a deliberate structure keeps the human inside the loop that builds the capacities, or whether the default, which is removal of that loop, is allowed to operate unopposed. The next section specifies that structure exactly, because it is the entire difference between the two outcomes.

6. The condition that decides which way a capacity moves

6.1 Two conditions, from the science of expertise

The deciding structure is not a matter of judgement or taste. It can be stated with precision, because the science of expertise already stated it, for a different purpose, decades before this question arose.

Kahneman and Klein, reconciling two research traditions that had been treated as opposed, set out the two conditions under which skilled intuition is genuine rather than illusory. A capacity to judge well in a domain develops, and is trustworthy, only when both of the following hold. First, the environment must be sufficiently regular to be predictable, so that there are real patterns available to be learned rather than noise that merely looks like pattern. Second, the person must have had the opportunity to learn those patterns through prolonged practice with feedback that is timely and informative enough to correct error. Where both conditions hold, expertise forms and the felt sense is reliable. Where either fails, what forms instead is confident intuition with no validity behind it, indistinguishable from the inside from the real thing.

These two conditions are the entire engine of this paper's central claim, because automation acts directly on both of them, and acts on them in either direction depending on how it is structured.

It is worth noting how unusual it is to have a precise, pre-existing, empirically grounded specification of the deciding condition before the question is asked. The Kahneman and Klein conditions were not developed to address automation. They were developed to settle a decades-old dispute between two schools of decision research, and they happen, when carried over, to give an exact answer to the question of which human capacities survive the automation of execution and under what circumstances. The paper is not constructing a criterion to fit a desired conclusion. It is borrowing one that was already established, for other reasons, by adversaries who had no stake in this argument, which is the strongest position from which a criterion can be applied.

6.2 How automation can satisfy both conditions, or destroy both

Automation can satisfy the first condition or destroy it. A system that behaves predictably, that signals what it is doing and why, that fails in legible and repeating ways, keeps the environment learnable, so a human working alongside it can build and maintain real patterns about when it is right, when it drifts, and what its characteristic errors look like. A system that behaves unpredictably, that changes its behaviour without signal, that fails in ways that do not repeat or that present identically to its successes, destroys the regularity, so no genuine expertise about it can form, and any felt sense the human develops about it is the invalid kind that feels exactly like the valid kind. The first condition is therefore not a property of the human. It is a property of how the automated environment was designed, and it is set by the designer whether or not the designer knows they are setting it. This connects directly to the trauma-informed and identity-layer arguments developed elsewhere in this library: a system that is unpredictable to the person who depends on it is not merely stressful, it is unlearnable, and an unlearnable environment is one in which no genuine human expertise about the system can ever form. Predictability is not, on this reading, a comfort. It is the precondition of the human keeping any calibrated relationship to the machine at all.

Automation can satisfy the second condition or destroy it, and this is the sharper of the two. The second condition requires feedback-bearing practice. The default behaviour of automation, the behaviour Bainbridge named, is to remove exactly the practice that the condition requires, by taking the routine work that was the substrate of the practice and returning to the human only the supervisory position, which carries no practice and generates no feedback because the human is no longer doing the thing, only watching its output. A human in the pure supervisory position is in the precise configuration in which the second condition fails: they have an environment but no practice in it, so their capacities do not form and the ones they had decay, while their confidence, anchored to familiarity rather than competence, does not move. A structure that keeps the human doing enough of the consequential work to generate real feedback, that routes hard cases to the human as practice rather than only as emergencies, that closes the loop between the human's judgement and its outcome so error is felt and corrected, satisfies the second condition and the capacities appreciate. A structure that seats the human above an output stream they no longer practise against fails it, and the capacities erode regardless of how able the human was when the structure was put in place.

There is a counterintuitive consequence here that is worth drawing out, because it overturns the most common design instinct. The instinct, when automating, is to give the human the easy cases and reserve the hard ones for the machine where the machine is strong, or, more often, to give the machine everything and reserve only the exceptions for the human. Both arrangements fail the second condition, and they fail it precisely where it matters. A human who only ever sees the exceptions has no practice on the ordinary cases against which the exceptions are exceptional, so they have lost the baseline that made the exception legible as one. A human who only ever sees the easy cases is practising, but on material that builds nothing, because easy cases generate no informative feedback: they are right by default and teach the judgement nothing. The arrangement that satisfies the condition is the one that feels least efficient in the short run: routing a deliberate, sustained share of the genuinely consequential and genuinely uncertain work to the human, not because the machine could not attempt it, but because that is the only material on which the human's judgement is maintained. The practice loop is not preserved by giving the human what is left over. It is preserved by giving the human, on purpose, work the machine could have done, because the point of the work is no longer only the output. It is the maintenance of the faculty that will be needed when the output cannot be trusted.

6.3 The condition restated as a single principle

The two conditions collapse into one principle once their common element is seen. A human capacity appreciates under automation when, and only when, the human remains inside a practice loop that is both learnable and feedback-bearing with respect to the consequential work. It depreciates when the human is moved outside that loop into supervision of an output they no longer practise against. The capability of the machine is not in this principle at all. A more capable machine that keeps the human in the loop builds the human's capacities faster, because it raises the quality of the work the human is practising judgement against. A less capable machine that removes the human from the loop erodes them anyway. The variable that determines the outcome is the structure of the loop, not the power of the model, which is why this is an architecture question and not a technology question, and why it is decided by deliberate design rather than by the pace of capability.

This is the precise point at which the paper parts company with both of the genres it disavowed in Section 1.4. Against the consoling forecast, it does not claim the human is safe; the human is safe only if the loop is kept open, and the default is that it is not. Against the deflationary certainty, it does not claim the capacities are doomed; they are doomed only if the loop is allowed to close, and whether it closes is a design decision, not a law of technological progress. The future the paper describes is not a prediction about what models will be able to do. It is a statement about what structure is chosen, by whom, and when, which is why it is the rare argument of this kind that terminates in something a person can actually build rather than only something they can brace for.

6.4 The meta-skill, and the honest expiry of its clearest example

There is a specific capacity that this principle elevates above the rest, and the clearest illustration of it has an honest expiry that the illustration must include, because using it without the expiry would be the kind of unsupported claim this library refuses.

When Garry Kasparov introduced advanced chess in the late 1990s, allowing a human and a machine to play as a team, the freestyle tournaments that followed in the mid-2000s produced a striking result. Well-coordinated human and machine teams, called centaurs, beat both the strongest unaided humans and the strongest unaided engines of that period, and the best centaurs were often not grandmasters but skilled operators who were expert not at chess but at directing and verifying the machine: knowing which of its lines to trust, when to override it, where its judgement was reliable and where it was not. That meta-skill, the disciplined direction and calibrated verification of a more capable instrument, is the practical name for the faculty this paper has been specifying.

The honest part, which any serious treatment must state, is that the chess example expired. By the late 2010s, engine strength advanced so far that the centaur's edge in chess narrowed to almost nothing, and a strong engine alone became, for that bounded and fully formalised game, as good as or better than the human and machine together. This is not a weakness in the argument. It is the argument made precisely. The centaur's edge persisted exactly as long as the meta-skill of direction and verification added value the machine could not supply itself, and it expired in the one domain, a closed game with a perfect evaluation function and complete information, where the conditions for the human contribution structurally dissolve. The lesson is not that human and machine teams always win. The lesson is that the meta-skill appreciates precisely in the domains the chess endgame is not: open, ill-structured, value-laden, incompletely specified, with no perfect evaluation function and no complete information, which is to say almost all consequential human work. The example's expiry marks the boundary of the claim rather than undermining it, and stating the expiry is what makes the surviving claim trustworthy.

7. The layer on which the condition is met or missed

7.1 The condition is not satisfied by individuals alone

Everything in Section 6 was stated as if the practice loop were something an individual could maintain by personal discipline. That framing is incomplete and, left uncorrected, would quietly weaken the argument, because the loop is not, in any institution of consequence, an individual artefact. It is a structural one.

The reason is that the loop has to persist across the discontinuities that individuals do not survive intact. People change roles, leave, forget, and are replaced. The work passes through many hands and many systems. The model underneath the work is rebuilt on a cadence faster than any individual's tenure. A practice loop maintained only in one person's habits is destroyed every time that person is replaced and degraded every time their attention is, which is to say it is not durable at the level the institution actually operates on. For the condition in Section 5 to hold at institutional scale, the structure that keeps humans inside the feedback-bearing loop has to be carried at a layer below any individual: a deliberate, maintained structure that decides what work humans keep doing rather than cede, what is routed to a person as practice rather than only as emergency, what is verified against what, what feedback is captured and returned so error is felt and corrected, and what the work is ultimately for, so the direction the capacities serve does not quietly become the system's instead of the person's.

7.2 The substrate that holds the loop open

That layer has a precise description, and it is the description toward which the whole paper has been moving. It is the carried-forward structure of an institution: the encoded decisions about what humans remain in the loop on, the accumulated and maintained record of how the institution's judgement was formed and corrected over time, the standing account of what the work is for and whom it is for, held below the level of any individual and persisting across every change of person and model. This is the substrate. Its function, in the terms of this paper, is exactly to keep the two conditions of Section 6 satisfied at institutional scale and across time: to keep the environment learnable by holding what was learned, and to keep the practice loop feedback-bearing by deciding, deliberately and continuously, what the human keeps doing and how the consequence of doing it returns to inform the next instance.

This reframes what the maintenance of such a substrate is. It is not documentation and it is not administration. It is the active construction of the only condition under which the emergent human capacities appreciate rather than erode. An institution that maintains it well is, whether or not it describes the work this way, continuously deciding to keep its people inside the loop that builds judgement. An institution that does not maintain it allows the default to operate, and the default, as Section 5 established with the full weight of the evidence, is erosion behind undiminished confidence.

7.3 Inside the question before it is asked

It is worth being exact about the timing, because the timing is the part most easily missed. The institutions that will hold the appreciating capacities a decade from now are not the ones that will decide, a decade from now, to start preserving the practice loop. The capacities do not regenerate on demand. A judgement faculty hollowed over years of pure supervision is not restored by a later resolution to think harder, because the practice that would have built it did not happen and the years in which it could have are not available again. The structure that keeps the loop open has to already exist while the capacities are still intact, because its entire function is to prevent a loss that is invisible until it is decisive and irreversible once it is.

This is the level at which Future Proof has been working. Not at the layer where execution is being automated and commoditised, which is where the field's effort and the field's anxiety are both concentrated, but at the layer underneath it: the substrate that keeps the human practice loop open across time, that holds what the institution has learned so the environment stays learnable, that decides deliberately what the people remain in the loop on so the capacities keep forming rather than draining behind a confidence that does not fall. The treatment of an AI-era trust standard, of an identity layer held over the orchestration layer, of certification and assurance not as products but as the roots of a system that is in the act of hardening, is, read through this paper, the treatment of precisely the layer on which the emergent human capacities are either compounded or lost. A reader does not need to take any position on Future Proof to follow the argument. But a reader who has followed it will recognise that the layer this paper has been describing is not hypothetical and not future. It is already being operated, quietly, by those who reached it before the question this paper asks became the question everyone is asking.

8. Implications

The argument is structural, so the implications are not advice. They are consequences, and the same structural fact lands differently depending on where the reader is standing.

8.1 For institutions

An institution that measures competence by output is now measuring nothing about its own durability, because output no longer implies the faculty that used to produce it. The cut arrow in Figure 1 is the operative fact: a bright artefact at the bottom of the stack is, after automation, evidence of the machine and evidence of nothing about the human faculty above it. An institution that continues to read competence off output will conclude, falsely and comfortably, that its capacities are intact long after the practice loop that maintained them has been quietly removed, because the output stayed bright the whole time and the confidence never fell.

The operative implication is that the practice loop must be treated as primary infrastructure, designed and maintained as deliberately as the systems that automate the execution, and on the same footing as cost and speed rather than after them. The decision about what work people keep doing is not a staffing decision and not a cost decision. It is the decision that determines whether the institution's judgement appreciates or erodes, and it is being made, by default and usually unexamined, every time a task is automated without a deliberate answer to the question of where the human practice now lives. Most institutions are making that decision by not noticing it is a decision, and the direction of the default is established.

There is a measurement implication that follows and is worth stating because it is actionable without being a prescription. If output no longer evidences the faculty, then an institution that wants to know the state of its own judgement cannot read it from deliverables and has to test it directly: by occasionally removing the automation and observing whether the judgement is still there, by tracking whether its people can still detect the smooth error rather than only produce the smooth output, by treating the capacity to override the machine correctly as a measured property rather than an assumed one. An institution that has never tested whether its people can still work without the system does not know whether they can, and the confidence inversion guarantees that asking them will not reveal it, because they will report that they can, and they will believe it, and the report will be uninformative for exactly the reason this paper has established.

8.2 For investors

The investable question in this era is not which entity has the most capable model or the most automated execution. Those are converging toward a utility and the advantage they confer is the kind that has to be re-won every cycle. The investable question is whether the entity has preserved, deliberately and structurally, the practice loop that keeps its human judgement appreciating, because that judgement is the part that does not commoditise and the part a better-funded competitor with a better model cannot reconstitute, since the missing input on the other side is accumulated feedback-bearing practice and that input is time, which capital cannot purchase.

This implies a different diligence. The durable question is not what an entity can produce today, because production is exactly the thing becoming abundant and cheap. It is whether the entity's people are still inside the loop that builds the upstream faculty, whether the structure that keeps them there is maintained below the level of any individual, and whether the entity can tell the difference between judgement that is sound and judgement that is merely confident, which is the same calibration question this paper has argued is now the load-bearing one. An entity with impressive output and a hollowed loop is a depreciating asset wearing an appreciating one's clothes, and the confidence inversion documented in the 2025 evidence is the precise reason that condition is hard to detect from the output and easy to detect only by looking at the structure.

8.3 For operators

For the operator working inside this, the implication is concrete and immediate, and it inverts the intuitive priority. The instinct under automation is to move as fast as possible to pure supervision, to get above the work and let the machine carry all of it, because that feels like the highest-leverage position and the one the trajectory rewards. The science says that is the precise configuration in which the second condition of Section 6 fails and the operator's own capacities begin to erode behind a confidence that will not warn them.

The reframing is that staying inside the loop on the consequential work is not a failure to scale. It is the maintenance of the only thing the operator owns that appreciates. The discipline is to automate the execution while deliberately keeping a feedback-bearing practice in the judgement, the sense making, the verification, and the direction, to route the hard case to oneself as practice and not only as emergency, and to close the loop between one's own judgement and its outcome so that error is felt and corrected rather than smoothed over by an output that always looks finished. An operator who treats this as overhead taken away from real work is, without noticing, declining to maintain the one capacity that does not depreciate. An operator who treats it as the real work is compounding it.

8.4 For the people inside these systems

There is a final reader: the person whose judgement, attention, and direction the system is operating alongside, and whose felt loss when their competent work is automated this paper has insisted is structural and real rather than nostalgic.

For that person the implication is protective and it is honest in both directions. It is honest that the specific competence automation depreciates is genuinely depreciated, and that the appreciation described here is at a different layer and is not automatic consolation for that loss. And it is protective in naming exactly what does appreciate and under exactly what condition, so the condition can be demanded rather than hoped for. The capacities that appreciate are the orienting faculty: deciding what is worth doing, reading what the situation actually is, choosing under uncertainty, feeling before articulating that something is wrong, and assigning every claim including the machine's the trust it has earned. They appreciate if and only if the person stays inside a learnable, feedback-bearing practice loop with respect to the consequential work. Whether that loop is kept open is not, for most people, theirs alone to decide. It is decided by the structure they work inside, which is why the structure, and not the model, is the thing that determines whether the intelligence era makes them more capable or quietly less, while telling them, the entire time, that they are fine.

9. Coda

The thing the field called competence was never the competence. It was the execution, which was competence's visible proxy, welded to it only because, for the whole of working history until now, the artefact could not be produced without the faculty that produced it. The weld is being cut. The artefact comes now from the side, from a machine, and the bright output at the bottom of the stack no longer implies anything at all about the faculty that used to stand above it.

What that faculty is can now, for the first time, be seen whole, because it is no longer hidden behind the act of executing. It is the orienting of a person and their instruments toward an end they have reason to want, under conditions that do not resolve themselves: direction, sense making, judgement, the trained felt sense that something is wrong before it can be said, and the calibration that assigns every claim the trust it has earned. These appreciate as execution becomes abundant, for the oldest reason in economics, because value moves to the adjacent thing that did not become cheap, and none of these became cheap, because none of them is what the machine produces.

But the appreciation is conditional, and the condition is the single idea to carry out of this paper. The same automation that makes these capacities more valuable also, by default and through a mechanism named in 1983 and confirmed in every decade since, erodes the practice loop that builds them, and it does so silently, because the confidence that masks the erosion does not fall with the competence it was supposed to track. A person being hollowed by this does not feel hollowed. They feel exactly as sure as they did before, and the output in front of them stays bright the entire time. Which is why the decisive work of this era is not building the machine that executes. It is maintaining, deliberately and below the level of any individual, the structure that keeps the human inside the loop that builds judgement, before the loop is removed by a default that announces nothing and is irreversible once it has run. The capacities are emergent. Whether they emerge or quietly drain is not decided by the machine. It is decided, now, by whoever is building the layer that holds the loop open, whether or not they know that is what they are deciding.

References and Notes

The following are real and verifiable sources. Where a precise figure circulating in 2025 and 2026 commentary could not be grounded to a primary source, or rests on self-report, the paper states the structural finding rather than the number, by editorial policy.

1. Ericsson, K. Anders, Ralf Th. Krampe, and Clemens Tesch-Romer. "The Role of Deliberate Practice in the Acquisition of Expert Performance." Psychological Review, vol. 100, no. 3, 1993, pp. 363 to 406. The foundational account of expertise as built through effortful, feedback-bearing practice rather than mere exposure or elapsed time.

1. Ericsson, Anders, and Robert Pool. Peak: Secrets from the New Science of Expertise. Boston: Houghton Mifflin Harcourt, 2016. The popular synthesis, and Ericsson's own correction of the misattributed ten thousand hour rule. The paper relies on the conservative core of this work, not the contested strong dose-response form.

1. Kahneman, Daniel, and Gary Klein. "Conditions for Intuitive Expertise: A Failure to Disagree." American Psychologist, vol. 64, no. 6, 2009, pp. 515 to 526. The two conditions for valid skilled intuition, environmental regularity and feedback-bearing practice, and the finding that subjective confidence is not a valid cue to accuracy. The hinge of the paper's central argument.

1. Tetlock, Philip E. Expert Political Judgment: How Good Is It? How Can We Know? Princeton: Princeton University Press, 2005. The forecasting tournaments, the near-chance accuracy of typical expert judgement, and the fox and hedgehog distinction after Isaiah Berlin.

1. Tetlock, Philip E., and Dan Gardner. Superforecasting: The Art and Science of Prediction. New York: Crown, 2015. The Good Judgment Project, calibration and resolution borrowed from meteorology, and the demonstration that judgement under uncertainty is a trainable and distributable skill.

1. Bainbridge, Lisanne. "Ironies of Automation." Automatica, vol. 19, no. 6, 1983, pp. 775 to 779. The mechanism by which automating the routine removes the practice that builds the competence the human needs for the non-routine, and then holds the human responsible for the failure it arranged.

1. Parasuraman, Raja, and Dietrich H. Manzey. "Complacency and Bias in Human Use of Automation: An Attentional Integration." Human Factors, vol. 52, no. 3, 2010, pp. 381 to 410. Automation complacency and automation bias as attentional phenomena occurring in experts as well as novices and not removed by training or instruction.

1. Parasuraman, Raja, and Victor Riley. "Humans and Automation: Use, Misuse, Disuse, Abuse." Human Factors, vol. 39, no. 2, 1997, pp. 230 to 253. The earlier framework on the patterns of human reliance on and rejection of automation.

1. Lee, Hao-Ping, et al. "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects from a Survey of Knowledge Workers." Microsoft Research and Carnegie Mellon University, 2025. Cited for the structural confidence inversion, not for its self-reported magnitudes, by editorial policy.

1. Gerlich, Michael. "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking." Societies, vol. 15, no. 1, article 6, 2025. Cited for the direction of the cognitive-offloading association, with its survey methodology stated openly, not for its precise effect sizes.

1. Polanyi, Michael. The Tacit Dimension. London: Routledge and Kegan Paul, 1966. The formulation that we can know more than we can tell, grounding the account of the felt sense and of sense making as partly uncodifiable.

1. Dreyfus, Hubert L., and Stuart E. Dreyfus. Mind Over Machine: The Power of Human Intuition and Expertise in the Era of the Computer. New York: Free Press, 1986. The novice-to-expert skill acquisition model and the account of expert situational perception. Cited for the skill model and the situated-coping thesis, not for strong impossibility claims.

1. Csikszentmihalyi, Mihaly. Flow: The Psychology of Optimal Experience. New York: Harper and Row, 1990. The autotelic experience and the dependence of optimal engagement on a skill exercised at its edge against real resistance. Cited as an interpretive frame.

1. Frankl, Viktor E. Man's Search for Meaning. Boston: Beacon Press, multiple editions, originally published 1946. The account of meaning as ensuing from dedication beyond the self rather than pursued directly. Cited as an interpretive frame.

1. MacIntyre, Alasdair. After Virtue: A Study in Moral Theory. Notre Dame: University of Notre Dame Press, 1981. The distinction between goods internal and external to a practice, used to specify why the loss of meaning under automation is structural rather than nostalgic.

1. World Economic Forum. The Future of Jobs Report 2025. Geneva: World Economic Forum, January 2025. Cited for the labour-demand corroboration that analytical thinking, resilience, creative thinking, systems thinking, and lifelong learning are the rising core skills.

1. Organisation for Economic Co-operation and Development. "Artificial Intelligence and the Changing Demand for Skills in the Labour Market." OECD, 2024, and the OECD programme on AI in work, innovation, productivity and skills, 2026. Cited for the structural finding that demand in highly AI-exposed occupations is shifting toward cognitive and non-routine skills.

1. The European Union Artificial Intelligence Act and the Digital Omnibus on AI. Referenced only for the structural point that accountability for AI-mediated work is hardening into law during this period. The precise applicability dates were, as of early 2026, the subject of ongoing legislative negotiation, with the Act fully applicable on 2 August 2026 as written and a proposed deferral of high-risk obligations under negotiation; the paper states the direction of the hardening rather than cementing a contested date, by editorial policy.

A note on the FP register. Future Proof Intelligence is referenced in this paper only as an existing body of practice operating at the layer that holds the human practice loop open across time, never as a product. No internal data, figures, named partners, or priced offerings appear, by editorial policy. The argument stands on its own sources and would stand without the reference.

Future Proof Intelligence. Research. No. XIII. MMXXVI.

Future Proof Intelligence . Research . No. XIII . MMXXVI