The Future of Talent

How a generation should be formed for work in environments where intelligent tools do much of the execution, and why the answer is structural rather than topical.

Future Proof Intelligence. Research. No. X. MMXXVI

Abstract

The question almost everyone is asking about the next generation and artificial intelligence is which skills to teach. This paper argues that the question is posed at the wrong level. The durable issue is not the content of a curriculum. It is the mechanism of formation itself. The science of expertise is unusually clear that judgement, taste, calibration, and the capacity to direct, verify, and take responsibility for a system are not transmitted by instruction or exposure. They are grown through effortful, structured, feedback rich practice on work that has real consequences. The same science is equally clear about which work that is: well specified, codified, consequence bearing tasks performed under feedback. That is precisely the band of work that intelligent tools automate first, and the early labour market signal of 2026 is the leading edge of its removal. The formation substrate and the automation target are therefore not adjacent. They are the same material. This is the formation paradox, and it is an identity, not a side effect. The paper sets out the four durable capacities with the calibration capacity as their governor, shows why generic skills training and topical AI courses do not reach the problem because the transfer literature says they cannot, and argues that serious education and early career formation must change structurally: what is assessed, what may be offloaded, and what a person is made answerable for. Formation, correctly understood, is an identity and continuity problem before it is a curriculum problem, and that is the layer this work is built underneath.

1. The question, posed at the wrong level

1.1 What is actually being asked

Every institution that touches the formation of young people is, in 2026, asking a version of one question. Schools ask which competencies to teach. Universities ask which programmes to redesign. Employers ask which graduates to hire and which skills to screen for. Governments ask which national strategy will keep a workforce employable. Underneath the variation, the question has a common shape: given that intelligent tools can now do a large and growing share of the execution that used to constitute entry level professional work, what should the next generation learn instead.

The question is reasonable and it is almost universally posed at the wrong level. It is posed as a question about content, the answer to which is a list: prompt engineering, data literacy, the ability to work alongside automated systems, durable human skills, critical thinking, creativity, adaptability. Lists of this kind are now produced in volume by ministries, consultancies, and the firms selling the tools. They are not wrong so much as they are answering a question the situation did not ask. The situation is not asking what to put on a syllabus. It is asking what happens to the mechanism by which a person becomes capable when the work that the mechanism runs on is taken away.

This paper is about that second question. It treats the future of talent not as a curriculum problem but as a formation problem, and it argues that the two are different in kind. A curriculum problem is solved by deciding what to teach. A formation problem is solved by understanding how a specific human capacity is actually grown, what conditions that growth requires, and what it costs when those conditions are removed. The science needed to answer the second question already exists. It has existed for decades. It has simply not been pointed at this situation, because the situation is new and the instinct under novelty is to reach for a list.

1.2 Why the level matters

The reason the level of the question matters is not pedantic. A list of skills, even a correct one, contains a hidden and false assumption: that the items on it can be taught the way content is taught, by being delivered, practised a little, and assessed. Some things can be formed that way. The capacities that this paper will argue actually matter cannot. They are not knowledge that can be delivered. They are dispositions and competences that are grown only under particular conditions, and the central, uncomfortable fact of the next decade is that the conditions under which they are grown are the same conditions that intelligent tools are removing.

If that is true, then the entire genre of response that consists of adding artificial intelligence to a curriculum, teaching students to use the tools well, and declaring the formation problem addressed is not merely insufficient. It is a category error. It treats a formation problem as a content problem and therefore cannot, even in principle, solve it, in the same way that adding a chapter on swimming to a textbook does not produce a person who can swim. The argument that follows is built to be judged on whether this claim is true, and the paper will spend its length establishing it from sources that are not in serious dispute.

1.3 What this paper is not

A short clarification of scope, because the topic attracts two adjacent genres this paper is not. It is not coding bootcamp futurism: the position that the answer is to teach everyone to build with the tools, that fluency with the orchestration layer is itself the durable skill. The paper will argue, on the evidence, close to the opposite. And it is not motivational: it does not conclude with an exhortation to be more human, more creative, more adaptable. Exhortation is the response of a system that has located its problem in the motives of individuals. This paper locates the problem in structure, and structural problems are not solved by encouragement. They are solved by changing the structure, which is harder, slower, and the only thing that works.

The paper is also a deliberate sibling to two others in this library. One examines the system that forms early career talent across borders and the ways it fails the person inside it. Another examines which human capacities appreciate rather than depreciate as machines absorb the tasks that used to define competence. This paper sits between them at a specific point: not the system that forms talent, and not the abstract question of what appreciates, but the concrete mechanism of formation itself, and what must change in education and early career so that the mechanism survives contact with abundant execution.

2. What an execution abundant environment actually is

2.1 The structure of the change, stated precisely

To reason about formation under automation it is necessary first to be precise about what is automating. The honest description is not that intelligent tools do everything, nor that they do nothing important. It is more specific and more consequential than either. The work that current systems perform most reliably is work that is well specified, codified, and decomposable into steps whose correctness can be checked against a clear standard. Drafting a document to a brief, producing a first version of an analysis from defined inputs, generating and adjusting routine code, summarising a body of material, vetting a dataset against rules, assembling a model from known components: this is the band of work where the tools are strongest and where their use is, by observation, most concentrated.

This is not a new structural insight. The task based literature on automation, developed over more than a decade before the current systems existed, established the principle that what automates first is what can be specified. The economists who built that literature, working on routine and codified task content, were describing the shape of the present before the present arrived. The current systems have not changed the shape. They have extended the boundary of what counts as specifiable far enough into cognitive work that the principle now reaches the desk of the junior analyst, the trainee, the graduate in their first year. The structure is old. Its reach is new.

The recent empirical mapping of where these tools are actually used, rather than where they could in theory be used, confirms the shape rather than complicating it. The publicly reported task level analyses of observed usage concentrate on exactly the codified, well specified category, and they distinguish, importantly, between use that automates a task entirely and use that augments a person doing it. The distinction matters for this paper because it is the difference between a tool that removes the formative work and a tool that changes how the formative work is done. The paper returns to that distinction, because almost everything turns on it.

2.2 The real signal of 2026

It is tempting, in a paper of this kind, to open the question of labour market effect with a single dramatic figure. The honest position is more useful and, here, more precise. As of 2026 there is a real and credible early signal, and it has a specific shape that matters more than any single number drawn from it.

The clearest public evidence comes from the analysis of high frequency payroll records covering millions of workers, conducted by researchers at a major economic research laboratory and reported under the title that compares early career workers to canaries in a coal mine. The finding, stated at the resolution at which it is solid, is directional and consistent across specifications even though the precise magnitude depends on the specification chosen. Early career workers, those in roughly the first years after entering the labour market, in the occupations most exposed to these tools, have experienced a relative decline in employment since the widespread adoption of generative systems in late 2022, while older workers in the same occupations have not, and in several measures have continued to grow. The effect appears only after the adoption point, is concentrated where the tools automate rather than augment, and is not confined to a single sector.

Read this signal correctly. The interesting thing about it is not its size, which will be argued about and revised, but its location. The effect lands first and hardest on the entrant. This is not the pattern of a technology that reduces total demand for a kind of work. It is the pattern of a technology that substitutes specifically for the work that entrants do, which is, definitionally, the codified well specified band described above. The corroborating employer survey evidence, the most cited of which reports that employers expect a large fraction of core skills to change within the decade and names analytical thinking as the most demanded core skill alongside a cluster of human competences, points the same way, with the standing caveat that stated employer demand measures what employers say they want and not what the work structurally requires. The two instruments, payroll microdata and employer survey, are weak in different directions and agree on direction. That agreement is the signal. Its meaning is the rest of this paper.

2.3 Why the location of the effect is the whole point

The location of the early effect, on the entrant rather than on the experienced practitioner, is not an incidental feature of the transition. It is the feature that makes the formation question urgent rather than merely interesting, and it is worth stating now even though the argument that establishes it does not arrive until section four.

If the work that is being substituted for is the work that entrants do, and if the work that entrants do is the work through which entrants become experienced practitioners, then a technology that removes entrant work is not only changing who is employed this year. It is interrupting the process by which this year's entrants would have become next decade's experienced practitioners. The first effect is visible on a one year horizon and shows up in payroll data. The second effect is invisible on any horizon shorter than the time it takes to make an expert, which is long, and it does not show up in payroll data at all because the people it concerns are still, in 2026, in school. The labour signal of 2026 is the early, measurable shadow of a deeper and currently unmeasured thing, and the deeper thing is a formation problem. Establishing why is the work of the next two sections.

3. How durable capacity is actually formed

3.1 The thing instruction cannot do

There is a large and unusually settled body of research on how human beings come to be good at difficult things. It spans the study of expertise, the cognitive science of learning, and the philosophy of knowledge, and across those fields it converges on a conclusion that is inconvenient for any institution that would prefer formation to be a matter of delivering content. The conclusion is that the capacities that distinguish a capable practitioner from a competent beginner are not transmitted by being told. They are grown by doing, under specific and demanding conditions, and the conditions are not optional.

The foundational work on the acquisition of expert performance established that what separates experts from everyone else, within domains where expertise is real, is not primarily talent or accumulated time but a particular kind of practice: effortful, deliberately structured to improve specific aspects of performance, conducted at the edge of current ability, and tightly coupled to feedback. There is a long and legitimate scholarly argument about how much of the variance in attained expertise this kind of practice explains, and this paper takes no position in that argument because it does not need to. The contested claim is the strong one, that structured practice explains nearly everything. The claim this paper relies on is the weak one, and the weak one is not contested by anyone serious: structured, effortful, feedback coupled practice is necessary for the development of high level skill, and passive exposure, instruction without practice, and practice without feedback on consequences do not produce it. That is the load bearing fact, and it is solid.

The model of skill acquisition that traces the path from novice to expert says the same thing from a different direction. The novice operates by following explicit rules because the novice has no stock of concrete experienced situations to draw on. The progression toward expertise is, precisely, the accumulation of experienced concrete situations until response becomes situational and fluent rather than rule bound. The path from rule follower to expert runs through situations. Remove the situations and the path does not become shorter. It ceases to exist, because there is nothing for the developing practitioner to accumulate.

3.2 The fluency trap

The cognitive science of learning adds a finding that is, for this paper, the single most important one, because it explains why the formation problem will be systematically misread by the people inside it until it is too late to act cheaply.

Research on the conditions that produce durable learning established that the conditions which make learning feel easy and fluent in the moment are frequently the conditions that produce the least durable retention and the weakest transfer, and that the conditions which make learning feel effortful and slow are frequently the ones that build capacity that lasts. The term of art for the second category is desirable difficulty: difficulty that is not an obstacle to learning but the mechanism of it. The corollary, sometimes called the fluency illusion, is that performance which feels smooth and effortless is an unreliable and often inverted proxy for learning. The learner who finds the task easy and the output good feels themselves becoming capable. The feeling is not evidence. It is, under many conditions, evidence of the opposite.

Hold this finding next to the defining property of a powerful execution tool. The entire value proposition of such a tool, correctly built, is that it makes the production of competent output feel fluent. The brief becomes a document, the question becomes an analysis, the requirement becomes working code, and the person experiences the smoothness as productivity and, fatally, as competence. The learning science says plainly what is happening: the conditions that produce the feeling of capability are being maximised at exactly the moment the conditions that produce actual capability are being removed. The fluency trap is not a risk at the margin of tool use. It is the central, structural feature of tool use, scaled from a single study session to an entire formative career, and it guarantees that the people undergoing the formation failure will be the last to perceive it, because the failure feels, from inside, like success.

3.3 Why generic skills do not save this

A natural and common response to the foregoing is to grant it and then reach for a familiar reassurance: that even if specific tasks automate, the answer is to form general capacities, critical thinking, problem solving, learning to learn, that transfer across domains and outlast any particular task. The reassurance is comforting and the research on transfer of learning is unkind to it.

The transfer literature, examined honestly, supports a narrow and sobering conclusion. Near transfer, the application of something learned to a closely similar situation, is real but limited. Far transfer, the application of a general capacity learned in one domain to a substantially different one, is rare, hard to produce, and routinely overclaimed. The strong position in that literature, that much of what is called transfer is an artefact of how similar the training and the test secretly were, has never been decisively refuted, and the practical upshot is robust even for those who find the strong position too strong: general thinking skills taught in the abstract do not reliably become general thinking skills exercised in the concrete. Capacity built in a domain tends to stay close to the domain it was built in.

This has a direct and underappreciated consequence for the formation question. It means the durable capacities this paper is about, judgement, taste, calibration, the ability to direct and verify and take responsibility, cannot be formed by a generic course in any of them, taught at one remove from real consequence, and then expected to transfer to the work. They are not generic. They are grown in contact with specific, consequential, domain embedded work, which is the very work that section two established automates first. The reassurance that we will simply teach the durable general skills instead does not survive contact with the literature on whether general skills, taught generally, are durable. They mostly are not. The capacities are real. They are just not transferable in the way the reassurance requires.

3.4 The part that cannot be told

There is a final element of the formation science that the rest of the argument rests on, and it is the oldest. A substantial part of real expertise is tacit: it is knowledge the expert possesses and uses but cannot fully articulate, the kind of knowing captured in the observation that we can know more than we can tell. Tacit knowledge is, by its nature, the part of competence that cannot be written into an instruction, extracted from artefacts, or delivered in a course, because if it could be made fully explicit it would no longer be the tacit part. It is transmitted, when it is transmitted at all, by participation: by working alongside, by attempting and failing and being corrected on consequences, by absorbing what is never said because it cannot be said.

This is the deepest reason the formation problem is not a content problem. The most valuable layer of a practitioner's competence is precisely the layer that no curriculum can carry, that no model can learn from finished outputs because it was never in the outputs, and that exists nowhere except inside people who acquired it by doing the work under the eyes of people who had it before them. A transition that removes the doing of the work does not merely make the explicit part of formation harder to deliver. It severs the only channel through which the tacit part has ever moved. That channel has no backup. It has never had a backup. Its existence has simply been assumed, the way the ground is assumed, by every profession that has ever reproduced itself.

4. The formation paradox

4.1 The identity, stated exactly

The three preceding observations can now be brought together, and when they are, they do not produce a worry. They produce an identity, in the strict sense: two things that looked separate are shown to be the same thing.

The science of expertise says durable capacity is formed by effortful, structured, feedback coupled practice on consequential, well specified, domain embedded work. The structure of automation says the work that is automated first is the consequential, well specified, codified, domain embedded work that can be checked against a standard. These are not two adjacent categories that happen to overlap at the edges. Read the two descriptions again with attention to their terms. They are descriptions of the same band of work. The work on which judgement is grown and the work that intelligent tools remove first are, to a degree that is not coincidental and not marginal, identical.

This is the formation paradox, and the word paradox is used precisely, not loosely. It is not the observation that automation has a downside for learners. It is the observation that the mechanism of formation and the mechanism of automation are competing for a single, finite, non substitutable input: real, consequential, feedback bearing repetitions performed by a developing practitioner. Every such repetition that a tool performs instead of the developing practitioner is not merely a task done faster. It is a unit of formation that did not happen and, by the science above, cannot be recovered later by instruction, because instruction was never how that unit was going to be delivered. The paradox is an identity because the thing being optimised away and the thing being grown are not two things. They are one thing, seen by two communities, the productivity community and the learning science community, who have not been in the same room.

4.2 Why this is invisible on every timeline that governs decisions

The formation paradox has a property that makes it almost perfectly designed to be ignored, and the property is worth naming because it is the reason action will be late unless the level of the question is raised deliberately, which is the work this paper exists to do.

The benefit of substituting a tool for the formative work is immediate, measurable, and accrues to whoever made the decision. Output is produced faster and at lower cost this quarter, and the line that records it moves in the right direction within the reporting period. The cost of that same substitution is the formation that did not occur, and that cost is invisible on every timeline shorter than the time it takes to make an expert. The person who would have been formed is, today, a student or a first year entrant. The absence of their formed judgement does not appear as a number anywhere. It appears, years later, as a cohort of senior practitioners who occupy senior roles without the depth those roles require, and by then the cause is a decade in the past, distributed across thousands of locally rational decisions, and attributable to no one.

This is the same structural shape that recurs throughout this library: a system that measures what is easy to count and is therefore blind to what matters, governed by metrics that can see occurrence and cost but cannot see formation. The labour signal of 2026 is the first thing about this that is countable, and it is countable only because it shows up as employment, which payroll systems record. The formation deficit does not show up in payroll systems. It shows up, eventually, in the quality of judgement available in the world, which no system records at all. A problem with this shape does not get solved by the actors inside it noticing, because the actors inside it are, each of them, behaving rationally against what they can see. It gets solved, if it gets solved, by something that changes what can be seen, which is the argument the paper builds toward.

4.3 Augmentation is not a loophole, unless it is built as one

There is one genuine point of leverage in the paradox and it is essential to state it precisely, because it is the difference between a counsel of despair and a design specification, and this paper is the second.

Section two drew a distinction between a tool that automates a task, performing it instead of the person, and a tool that augments a person performing it, changing how they do it. The formation paradox bites with full force only on automation, because automation removes the repetition entirely. Augmentation does not have to remove the repetition. In principle, a tool can be built and used so that the developing practitioner still performs the formative cognitive work, still struggles at the edge of ability, still receives feedback on consequences, while the tool removes only the part of the labour that is not formative. Whether augmentation actually does this is not a property of the technology. It is a property of how the technology is designed, deployed, and required to be used, which means it is a choice, and a contestable one.

This is the entire hinge of the practical argument. The fluency trap of section three guarantees that the path of least resistance, for both the learner and the institution, is to let the tool do the formative work, because that feels the most productive and the formation deficit it creates is invisible. Augmentation that preserves formation is therefore not the default. It is the harder path, it has to be deliberately constructed, and it runs directly against the in the moment incentives of everyone involved. The reform this paper argues for is, at its core, the deliberate construction and protection of the harder path: building the tools and the formative settings so that the developing practitioner is required to do the work that forms them, and required to be answerable for it, even though, and precisely because, a tool could have done it for them more fluently. Everything in the implications follows from this single point.

5. The four durable capacities

5.1 Why a list, after arguing against lists

This paper opened by objecting to the genre of skills lists. It now offers four capacities, which requires an account of why this is not the thing it criticised. The objection in section one was not to naming what matters. It was to treating what matters as content that can be delivered and to assuming, falsely, that the items transfer when taught generically. The four capacities below are named not as a syllabus but as a description of what the formative work, when it is preserved, actually forms. They are the output of the mechanism, not the input to a course. Each is defined rigorously, each is shown to be ungrowable by instruction alone, and one of them is shown to govern the other three.

5.2 Judgement

Judgement is the capacity to reach a defensible decision under irreducible uncertainty, where the information is incomplete, the options are not cleanly comparable, and no rule fully determines the answer. It is distinct from knowledge, which is what is known, and from analysis, which is what can be computed from what is known. Judgement is what is required precisely at the point where knowledge runs out and analysis cannot close the gap, which is the point that defines consequential professional work and the point at which a system that produces fluent answers is most dangerous, because the fluency of an answer is uncorrelated with the difficulty of the judgement it conceals.

The reason judgement cannot be delivered as content is structural and was established in section three. It is grown by making decisions under uncertainty and being held to their consequences, accumulating, over many such decisions, a situated sense of which considerations matter and how they trade off, much of it tacit. A model can produce a decision. It cannot, by producing it, grow the user's judgement, any more than watching a strong player makes one strong. If anything it does the reverse, because it removes the occasions on which the user would have had to decide, which were the occasions on which judgement would have formed.

5.3 Taste

Taste is the capacity to discriminate quality where quality is not reducible to a measurable specification: to know that this analysis, though it passes every check, is shallow; that this design, though it meets the brief, is wrong; that this argument, though valid, does not matter. Taste is judgement applied to quality rather than to decision, and it has the same structure: it is built by repeated exposure to the difference between good and excellent under conditions where the difference has consequences and is fed back, and it is largely tacit, which is why it is so hard to specify and so easy to dismiss as subjective by those who do not have it.

Taste becomes more important, not less, as execution becomes abundant, and the reason is exact. When producing competent output is cheap and universal, competent output ceases to be a differentiator, and the scarce thing becomes the discrimination of which competent output is actually good, which problem is actually worth solving, and which fluent answer is actually empty. The world about to be produced is one in which everyone can generate the competent and almost no one can reliably tell the excellent from the merely competent, because that discrimination was formed by exactly the practice the tools now absorb. Taste is the capacity that the abundance of execution makes simultaneously more valuable and harder to acquire, which is the formation paradox specialised to quality.

5.4 Calibration, the master capacity

Calibration is the accurate internal estimate of the reliability of a judgement: one's own, and now, decisively, a machine's. A well calibrated practitioner knows the boundary of their competence, knows how much to trust a given conclusion, and adjusts confidence to the actual evidence rather than to the coherence of the story available. The rigorous account of when this is possible at all establishes two conditions: the environment must be regular enough that valid cues exist, and the person must have had sufficient opportunity to learn those cues through feedback. The same account establishes the failure mode with unusual clarity: confidence tracks the coherence of the story a person can construct, not its validity, and experts characteristically do not know where the boundary of their own expertise lies. Confidence is not a measure of correctness. It is a measure of narrative fluency, which is precisely what a powerful generative system maximises.

This paper claims calibration is the master capacity, and the claim is structural, not rhetorical. Judgement without calibration is decision without knowledge of when to trust the decision, which is not judgement but its dangerous imitation. Taste without calibration is preference mistaken for discernment. And in an environment where the binding new skill is working with systems that produce confident, fluent, frequently wrong output, the capacity that governs whether any of the others can be safely exercised is the capacity to know how much to trust an answer, including, especially, the machine's. The new illiteracy of the intelligence era is not the inability to use the tools. It is the inability to calibrate them: to know, for this output, in this context, how much weight it can bear. Verification is calibration made operational, the disciplined act of checking an output rather than accepting its fluency. Responsibility is calibration made binding, the willingness to be the party who answers for the result. The two capacities most casually invoked in discussions of working with AI, verifying its output and owning the outcome, are not separate skills to be added to a list. They are calibration, instantiated as an action and as an obligation. This is why calibration governs the set.

5.5 The capacity to direct, and why it is not prompting

The remaining capacity is the ability to direct a system: to specify an end the system does not itself hold, decompose it into work the system can do, and integrate what comes back into a coherent result for which a person remains answerable. It is tempting and wrong to call this prompting. Prompting is the surface. Direction is the capacity beneath it, and the difference is the difference between knowing the syntax of an instruction and knowing what to instruct, why, to what standard, and how to tell whether the result is right, which is judgement, taste, and calibration applied to a system rather than to a task.

This matters for the formation argument because it disposes of the bootcamp futurist position directly. If the durable skill were operating the tools, formation would be easy: teach the interface, which is learnable in an afternoon and obsolete in a year. The durable capacity is not operating the tool. It is possessing the judgement, taste, and calibration required to direct it well and to know when it has failed, and those are exactly the capacities that, by the formation paradox, the tool's own competence makes harder to grow. The person who can direct intelligent systems to consequential effect is not the person who learned the tool. It is the person who was formed, by consequential work, into the judgement that knows what to ask of it and the calibration that knows whether to believe it. Direction is not a fifth thing. It is the first three, pointed at a system, with responsibility attached. There is no shortcut to it that does not pass through the formative work the tools are removing, which is, once more, the paradox, now stated as the refutation of the idea that the tools are their own remedy.

6. What serious formation must actually change

6.1 The reform is structural, not topical

If the argument to this point holds, the reform that follows is not the one currently being implemented almost everywhere, which is topical: add artificial intelligence to the curriculum, teach the tools, update the skills list. That reform is a category error for the reason established in section one, and the formation paradox shows why it cannot work even when done well, because the problem is not the absence of a topic. It is the removal of a mechanism. The reform that follows from the paradox is structural, and structural reform changes not what is taught but what is required, what is assessed, what may be offloaded, and what a person is made answerable for. Four changes follow directly, and they are demanding, which is the point. A reform to a formation problem that is not demanding is not addressing the formation problem.

6.2 Re introduce desirable difficulty deliberately, because it will not return on its own

The first change follows from the fluency trap. Once a tool can make the formative work feel effortless, the difficulty that the learning science says is the mechanism of durable learning does not merely become optional. It becomes something the entire incentive field of the learner and the institution actively removes, because removing it feels like progress and its absence is invisible until it is expensive. Therefore the difficulty has to be deliberately reintroduced and structurally protected, which means designating, within formation, zones in which the developing practitioner does the formative cognitive work without the tool, not as nostalgia and not as a ban on the technology, but for the same reason a pilot is still required to fly the aircraft by hand: because the capacity has to exist in the person, independent of the system, for the moments the system cannot be trusted, which are exactly the moments calibration says matter most.

This is more precise than a slogan about screen time or a prohibition. It is the claim that formation must contain a protected core of consequential work that the learner performs and is answerable for, with the tool deliberately withheld, and that what is protected is not arbitrary but specifically the work on which judgement, taste, and calibration are grown. Everywhere else in the workflow the tool can and should be used. The protected core is small, deliberate, and load bearing, and the act of designating it is the single most consequential decision a serious institution of formation will make this decade. Most are not making it. They are doing the opposite, integrating the tool everywhere because integration is what the moment rewards, which the formation paradox predicts will feel like modernisation and function as decay.

6.3 Assess the process, not the artefact, because the artefact no longer carries information

The second change follows from the collapse of the artefact as evidence. Education has, for its entire history, inferred formation from artefacts: the essay, the problem set, the analysis, the project. The inference worked because producing the artefact required the formation, so the artefact was a proxy for it. The execution tools sever that link completely. The artefact can now be produced without the formation, which means the artefact no longer carries information about whether the formation occurred, and an assessment system that continues to grade artefacts is, after this point, measuring nothing it intends to measure and certifying a thing it can no longer detect.

The structural response is to move the locus of assessment from the artefact to the process and the defence of it: to require the developing practitioner to show the work, to account for why an output is correct or where it fails, to defend a decision under questioning, to demonstrate the judgement rather than submit its residue. This is not a tooling preference. It is the only form of assessment that remains valid once the artefact is decoupled from the formation, because the process and its live defence are the parts a tool cannot perform on the learner's behalf without the absence becoming immediately visible, which is precisely the property the artefact has lost. The institutions that grasp this will assess differently and certify something real. The institutions that do not will continue to grade artefacts and will, within a decade, be issuing credentials that certify nothing, a fact that will be discovered the way the formation deficit is discovered: late, at scale, and by the people who relied on the credential.

6.4 Make responsibility a designed feature of formation, not an afterthought of employment

The third change follows from calibration being the master capacity and from responsibility being calibration made binding. Responsibility, in almost all current formation, is something that arrives after formation, in employment, when consequences first attach to a person's decisions. This sequencing was tolerable when the early career years supplied a long, graduated apprenticeship in which consequence increased slowly under supervision. The automation of entry level work, evidenced by the labour signal of section two, compresses or removes that apprenticeship, which means the gradual on ramp to responsibility is being removed at the same time as the work that used to constitute it. If responsibility is not designed into formation deliberately, the next generation will encounter genuine answerability for the first time at a seniority where the cost of immature judgement is high and the structure that used to grow it slowly is gone.

The structural change is to attach real, scaled consequence to decisions during formation rather than only after it: to construct formative settings in which the developing practitioner is genuinely answerable, within bounded stakes, for judgements they made and must defend, including judgements about when to trust a system and when to override it. This is the deliberate construction of the apprenticeship that the labour market is ceasing to provide, relocated upstream into formation because the place it used to live is being removed. It is demanding to build, because real consequence is harder to construct in a formative setting than a grade, and it is the only thing that grows the capacity, because the capacity is, by definition, the capacity to be answerable, and answerability cannot be simulated by an assessment that carries no weight.

6.5 Treat the literacy obligation as a floor and a temptation, never as the answer

The fourth change is a caution, and it is pointed at the most likely failure mode of the next three years. As of 2026 there is, in the European regulatory context, a live legal obligation that organisations ensure a sufficient level of artificial intelligence literacy among their staff and those operating systems on their behalf, an obligation in application since early 2025 with its enforcement architecture phasing through 2026. The obligation is correct, and it is necessary, and it is, for the formation question, a floor and a temptation in equal measure. It is a floor because a population that cannot use these systems competently is exposed. It is a temptation because compliance with a literacy obligation is satisfiable by exactly the topical training, teach the tools, the skills list, that the formation paradox shows cannot reach the actual problem, and a compliant programme will produce documentation that the obligation has been met while the formation deficit continues underneath it, unmeasured, because nothing in the obligation measures it.

This is the regulatory turn meeting the measurement trap, a pattern this library has described elsewhere and which recurs here in its sharpest form. A literacy requirement governs the floor: it ensures people can operate the systems. It does not, and by its nature cannot, govern the substance: whether the people operating the systems have the judgement, taste, and calibration to operate them well and to know when not to trust them. A formation regime that satisfies the obligation and stops has satisfied the part that is countable and ignored the part that matters, and it will record the first as success precisely because it cannot see the second at all. The obligation should be met. It should never be mistaken for having addressed the question this paper is about.

7. The missing layer

7.1 What the diagnosis converges on

Everything to this point has been a structural argument with no reference to any particular actor. It is worth saying plainly what the argument converges on, because it converges on a single thing and the single thing is not a curriculum.

Formation, correctly understood, is the slow accumulation, by a developing person, of judgement, taste, and calibration through consequential work for which they were answerable, much of it tacit, none of it transmissible by instruction alone. The intelligence era removes the consequential work that grew it and replaces the formative artefact with one a tool can produce, which severs the proxy by which formation was inferred and the channel through which the tacit part moved. The reforms in section six all have one property in common when looked at from the right distance. Each is an attempt to make formation visible, attributable, and continuous when the structures that used to carry it invisibly have been removed: to make the protected work and its defence into something the system can see, to make the process rather than the vanished artefact the thing of record, to make responsibility a tracked and binding feature rather than an evaporating one, to refuse to let a compliance record stand in for a continuity of actual formed judgement.

Read together, those are not four pedagogical preferences. They are four requirements for a single missing thing: a layer that holds, continuously and verifiably, who a developing person actually is as a maker of judgements, what they actually decided and were answerable for, and how that has accumulated over time, in a form that does not evaporate when an artefact is produced by a tool and does not reduce to a credential that certifies nothing. Formation has always depended on such a layer. It was simply carried, invisibly and unreliably, by apprenticeship, by the memory of the people who trained the next ones, by the tacit transmission that no one wrote down because it was assumed to be there. The intelligence era removes the carrier. It does not remove the need. It makes the need explicit for the first time, by taking away the thing that was meeting it without anyone having to name it.

7.2 Formation is an identity and continuity problem

This is the load bearing reframe of the paper and it should be stated without hedging. The future of talent is not, at its root, a question of what to teach. It is a question of whether there is a layer beneath formation that remembers what a person actually became, carries it continuously across the discontinuities of education and early career, and holds it as something true and theirs rather than something inferred from an artefact a machine could have produced. The skills question is real but downstream. Upstream of it is an identity and continuity question, and the identity and continuity question is the one the intelligence era has made unavoidable, because it has removed every informal mechanism that used to answer it without being asked.

This is the same structural conclusion that this library reaches from the direction of the cross border early career system, which fails its people through statelessness, an inability to remember how anyone actually behaved or what anyone actually became across the handoffs. It is the same conclusion the library reaches from the direction of the abundance of intelligence, where the durable layer is not the orchestration, which is becoming cheap, but the identity layer above it, which remains scarce. The future of talent is where those two arrive at the same place from the third direction, the human one. The orchestration of work is being automated and will be abundant. What stays scarce, and what therefore has to be deliberately built, is the layer that holds who a person is as a bearer of judgement, continuously, accountably, across every discontinuity, so that formation has somewhere to accumulate and someone to be attributed to. Without that layer, the most thoroughly resourced curriculum in the world forms judgement that the world has no way to see, trust, or carry, which is to say it forms it into an artefact and loses it the way the system loses everything it cannot remember.

7.3 Why this is a substrate and not a programme

The natural misreading of section 7.2 is to imagine a credential, a portfolio platform, a better transcript: a product that sits on top of education and competes inside it. That is the wrong mental model and correcting it is the precise point at which this paper says what it has been pointing at. A credential sits on top of formation and certifies its residue, and the entire argument of section 6.3 is that the residue no longer carries the information, so a better certificate of a hollow artefact is not a solution. It is the problem with a cleaner cover. The thing the diagnosis points at is underneath formation, not on top of it: the continuous, verifiable holding of who a person actually is as a maker of owned, calibrated judgements, the layer the curriculum and the credential and the employer all stand on and all currently assume without anyone having built it on purpose.

That is the register in which Future Proof should be understood here, and the only register in which it appears in this paper. The thesis running beneath this library is that the durable layer of the intelligence era is not the doing, which is becoming abundant, but identity, trust, and continuity of meaning across systems, which remains scarce. The formation of a generation is where that thesis stops being abstract and becomes a specific seventeen year old whose judgement is either being grown and held as theirs, continuously and accountably, or being substituted away one fluent output at a time and lost the moment it is produced. The work of building the layer that holds who a person is as they are formed, so that formation has somewhere to accumulate rather than nowhere to go, is not waiting for a paper to propose it. The paper exists because the building came first, and a serious foundation documents the ground it already stands on. Future Proof is not a programme inside formation. It is the layer underneath the question this paper has been asking, present already, the way ground is present, assumed rather than announced.

8. Implications

This section is divided by reader, because the same structural fact lands differently depending on where one stands. None of what follows is a pitch. It is the diagnosis applied.

8.1 For institutions

You are the layer with the clearest duty to form people and, increasingly, the least defensible instruments for proving you have. The artefact you have certified for your entire history no longer carries the information it used to carry, which means a credential that grades artefacts now certifies something it can no longer detect, and the discovery of this will not be gradual. The implication is not to add a course on the tools. It is that an institution's duty to form judgement has become, in practice, a problem of designating and protecting the formative core, assessing process rather than residue, and treating formation as something to be continuously held and attributed rather than inferred once and forgotten. The institutions that hold their people well in the next decade will be the ones that grasp that formation is an identity and continuity problem before it is a curriculum problem, and that the substrate which carries that continuity is something to stand on rather than something each institution rebuilds alone and badly.

8.2 For investors

The investable surface of education is usually read as content and delivery: better courses, better platforms, the tools themselves. This paper's structural reading is that the durable position is none of those. Content is abundant and the tools commoditise by construction. The scarce and appreciating layer is the one that holds, continuously and verifiably, who a person actually became as a bearer of judgement, because that layer compounds with time and use while everything above it depreciates. The relevant question for capital is not which curriculum or which tool wins. It is which layer the formation of a generation cannot, in a decade, function without once the informal carriers are gone, and who was already operating at that layer before the standard for it set. Value built on the abundance of content or the novelty of a tool is value exposed to the thing this paper describes. Value built on the layer beneath formation is value positioned where the scarcity is moving, not where it is leaving.

8.3 For operators

If you run a team that takes in early career people, you are making, every quarter, the locally rational decision the formation paradox describes: let the tool do the codified work because it is faster and the cost is invisible. You cannot solve this by caring more, and you should not try, because exhortation does not change a structure. What changes it is deliberately constructing the harder path: requiring the developing practitioner to do the formative work and be answerable for it even where a tool could have produced it more fluently, building augmentation that preserves the repetition rather than removing it, and treating the record of what your people actually decided and owned as a continuous asset rather than something that evaporates with each finished artefact. The operators who compound advantage will be the ones who understood that their pipeline of senior judgement is being severed upstream, on a timeline their quarterly metrics cannot see, and who chose to operate on a substrate that makes formation visible and attributable instead of letting it disappear into output.

8.4 For the people inside these systems

This is the reader the system is supposed to be for and structurally is failing first, because the labour signal of 2026 lands on the entrant. The most important thing this paper can say to a person being formed now is that the fluency they feel when a tool produces their work is not evidence that they are becoming capable, and that the learning science says it is frequently evidence of the opposite. That is not a reason for fear and it is not a reason to refuse the tools, which would be its own kind of unpreparedness. It is a reason to seek out, deliberately and against the path of least resistance, the consequential work the tool could have done for you, and to do it, and to be answerable for it, because that is the only thing that forms the judgement the same tools will then make scarce and valuable. And it is a reason to insist that what you actually became, the judgement you grew and owned, is held somewhere as true and yours, rather than lost the moment an output is produced. That layer should simply be there, the way ground is. Where it is not yet there, its absence is being priced and paid, and the people paying it are the ones being formed right now.

8.5 For the system as a whole

The four readings converge on one observation worth stating plainly. Every actor in the formation system is, in 2026, behaving rationally against what it can measure, and what it can measure is occurrence, cost, and compliance, not formation. The institution certifies the artefact because the artefact is what it can grade. The operator substitutes the tool because the saving is what it can see. The learner accepts the fluency because the fluency is what feels like progress. The regulator measures literacy because literacy is what an obligation can specify. Each is locally correct and the sum is a system that produces the measurable shell of a formed generation while being structurally blind to whether anyone was formed inside it. This is not solved by better motives. It is solved by a layer that makes formation itself visible, attributable, and continuous, so that the right thing becomes the measured thing and the actors can finally optimise for what matters because, for the first time, the system can see it. The test of whether a proposed layer is real infrastructure rather than another product competing inside the system is exactly that: real infrastructure makes the formation of judgement the thing the system can see and therefore the thing it rewards, without requiring anyone to be better than the structure lets them be. That is the standard against which the missing layer should be judged, and the standard the substrate beneath this question is built to meet.

Coda

There is a seventeen year old somewhere this morning who will spend the next decade being formed for work, and the system around her has, by every metric it can read, never been better resourced to do it. There is more content available to her than to any cohort in history, free and instant. There are tools that will produce, from a sentence of instruction, work that would have taken a competent adult a day. By every measure the system can see, she is being prepared for the future faster and more thoroughly than anyone before her.

The system cannot see the only thing that determines whether she is actually being formed, which is whether she is doing the consequential, effortful, answerable work on which judgement is grown, or whether that work is being done for her, fluently, one output at a time, while she experiences the fluency as competence because the learning science guarantees she will. That is the whole argument of this paper compressed into one person. The future of talent is not a question about what to teach her. It is a question about whether the work that would have formed her is being quietly removed in exchange for output, and whether anything is holding what she actually becomes, continuously and as hers, or letting it evaporate the moment a machine produces the artefact that used to prove it.

Build the layer that holds who she is becoming, and protect the work that forms her even though a tool could do it faster, and she is formed, and the world can see it, and trust it, and carry it. Do not, and the system will be, a decade from now, a more efficient version of exactly this: excellent at producing the measurable evidence of a prepared generation, structurally indifferent to whether anyone was formed underneath it, and unable to tell the difference precisely when the difference is everything. The work that forms judgement is being substituted away now, fluently, on a timeline the metrics cannot see. What gets built underneath that, while the ground is still soft, is the only part of this that is still ours to decide.

References and Notes

The following sources are real and publicly verifiable. Where this paper states a structural truth rather than a precise figure, it does so deliberately, because the precise figure either does not exist at the necessary resolution or is not solid enough to cement.

1. OECD. The Future of Education and Skills: Education 2030, The Future We Want (OECD Publishing, 2018); and the OECD Learning Compass 2030 and Future of Education and Skills 2030 concept note series (OECD Publishing, 2019), including the three transformative competencies (creating new value, reconciling tensions and dilemmas, taking responsibility), the Anticipation, Action, Reflection cycle, and the definition of student agency as the capacity to set a goal, reflect, and act responsibly to effect change. (oecd.org, project Future of Education and Skills 2030)

1. Ericsson, K. A., Krampe, R. T., and Tesch Romer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363 to 406. And Ericsson, K. A. et al. (eds), The Cambridge Handbook of Expertise and Expert Performance. Cited only for the uncontested core: that structured, effortful, feedback coupled practice is necessary for the development of high level skill, and that exposure and instruction without it are not sufficient. The stronger claim that such practice explains nearly all attained variance is contested in the literature (see Macnamara, Hambrick, and Oswald, and the associated debate in Frontiers in Psychology, 2019) and this paper does not rely on it.

1. Kahneman, D., and Klein, G. (2009). Conditions for intuitive expertise: a failure to disagree. American Psychologist, 64(6), 515 to 526. On the two conditions for valid intuitive skill (a sufficiently regular environment and adequate opportunity to learn it through feedback), on confidence tracking the coherence of the available story rather than its validity, and on experts not reliably knowing the boundary of their own expertise.

1. Detterman, D. K. (1993). The case for the prosecution: transfer as an epiphenomenon. In Detterman, D. K., and Sternberg, R. J. (eds), Transfer on Trial: Intelligence, Cognition, and Instruction. Norwood, NJ: Ablex. With Barnett, S. M., and Ceci, S. J. (2002), When and where do we apply what we learn? A taxonomy for far transfer, Psychological Bulletin, 128(4), 612 to 637. On the rarity and difficulty of far transfer and the limits of generically taught general skills.

1. Bjork, R. A., and Bjork, E. L. (2011). Making things hard on yourself, but in a good way: creating desirable difficulties to enhance learning. In Gernsbacher, M. A. et al. (eds), Psychology and the Real World. New York: Worth. On desirable difficulties as the mechanism rather than the obstacle of durable learning, and on the fluency illusion, the unreliability of in the moment ease as a proxy for learning. Companion: Roediger, H. L., and Karpicke, J. D. (2006), Test enhanced learning, Psychological Science, 17(3), 249 to 255.

1. Brynjolfsson, E., Chandar, B., and Chen, R. (2025). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab. On the relative decline in employment for early career workers in the most AI exposed occupations following the late 2022 adoption of generative systems, with older workers in the same occupations not showing the decline. Reported here as direction and order of magnitude with the specification dependence noted, not as a single cemented figure. (digitaleconomy.stanford.edu)

1. Anthropic. The Anthropic Economic Index and associated research on AI usage mapped to occupational tasks and on labor market impacts (2025 to 2026). On the concentration of observed AI use in codified, well specified tasks and the distinction between automating and augmenting use. Reported structurally rather than as precise universal percentages. (anthropic.com)

1. World Economic Forum. The Future of Jobs Report 2025 (Geneva, January 2025). On employers expecting a large fraction of core skills to change by 2030 and on analytical thinking as the most demanded core skill alongside resilience, flexibility, leadership, and curiosity. Treated as corroborating employer signal with the caveat that stated employer demand is a weak instrument for what work structurally requires. (weforum.org)

1. Polanyi, M. (1966). The Tacit Dimension. London: Routledge and Kegan Paul. On tacit knowledge, the principle that we can know more than we can tell, and the consequence that the most valuable layer of expertise is the one that cannot be made fully explicit and is transmitted only through participation.

1. Dreyfus, H. L., and Dreyfus, S. E. (1986). Mind over Machine: The Power of Human Intuition and Expertise in the Era of the Computer. New York: Free Press. On the five stage progression from rule following novice to intuitive expert and its dependence on the accumulation of experienced concrete situations.

1. European Union. Regulation (EU) 2024/1689 (the Artificial Intelligence Act), Article 4 on AI literacy, in application since 2 February 2025, with the enforcement and governance architecture phasing through 2026. On the legal obligation that providers and deployers ensure a sufficient level of AI literacy among staff and others operating systems on their behalf. (eur-lex.europa.eu, digital-strategy.ec.europa.eu)

1. Autor, D., Levy, F., and Murnane, R. J. (2003), The skill content of recent technological change, Quarterly Journal of Economics; and Frey, C. B., and Osborne, M. A. (2013, published 2017), The future of employment, Technological Forecasting and Social Change. Cited for the established structural principle that routine, codified, well specified task content is what automation reaches first.

Note on figures. The labour market evidence in section two is reported as direction and order of magnitude, attributed to its dataset and authors, because the precise magnitude of the early effect is specification dependent and still being revised, and cementing a single number would be less honest than stating the structural finding, which is solid: the early employment effect of generative systems falls disproportionately on early career workers in the most exposed, most codified work.

Note on the expertise literature. This paper deliberately relies only on the uncontested core of the deliberate practice research and takes no position in the active scholarly dispute about how much of attained expertise structured practice explains. Nothing in the argument depends on the contested claim, and the argument is stronger for not needing it.

Future Proof Intelligence. Research. No. X. MMXXVI.

Future Proof Intelligence . Research . No. X . MMXXVI