The future of human learning. Specialist, operator or ritualist?

The question is straightforward. If AI knows more than you in any reasonably codified domain it touches, what makes sense to learn? Three paths are already open and coexist in varying proportion across the professional population of 2026. First: be a specialist on a frontier AI doesn't reach well yet. Second: be an operator of AI, without yourself knowing what the AI does. Third: accept that learning loses its pragmatic function and becomes identity, ritual, an exercise of personal constitution. All three are already happening. None is the correct answer. They're the options available within a framework nobody negotiated.

Path 1: specialist on a moving frontier

The idea is recognizable. If AI covers the centre of codified knowledge, the professional survives by specializing in what AI doesn't cover well yet. The problem is that the frontier moves. What in January 2024 was a frontier —realistic image generation, reasonably complex code, step-by-step mathematical reasoning— was standard capability by May 2025. Every six to twelve months, the frontier moves a few kilometres further into territory that used to be human.

The specializations that hold up, according to the data available as of 2026, share at least one of these properties:

• Non-trivial physical domain. Fine manual surgery, hands-on physiotherapy, dentistry, specialized manual crafts, handling materials with tactile feedback. AI is progressing in robotics, but body-machine integration remains a practical barrier for many years.
• Human relationship where presence matters. Deep psychotherapy, palliative care, conflict mediation, early-childhood teaching. There are AI versions of these services, but the fraction of users who prefer —and pay for— human presence is robust.
• Concentrated legal responsibility. Surgeon, judge, forensic doctor, notary. The human signature has specific legal value the synthetic signature doesn't assume. As long as regulation doesn't change, the function stays shielded.
• Rare, high-value microspecialty. Subdisciplines with little demand but critical —pathologists of rare diseases, specialists in legacy software systems, translators of minority languages for international courts. The corpus to train AI in these tasks is insufficient and AI arrives late or never.

Newport, in So Good They Can't Ignore You (Grand Central, 2012), argued —before the generative AI era— that «become an expert in something specific» was the most reliable strategy for building professional capital. His thesis ages well in general; but it needs an appendix: the «something specific» must be selected by looking at where the AI frontier is most stable, not where the field is most glamorous.

This path has a cost. Specializing in a microfrontier takes many years of investment. Whoever bets on a microfrontier bets that the frontier won't close during those years of investment. The bet is reasonable, but not safe. The strategy works for a fraction of the professional population, not for everyone.

Path 2: operator of AI without knowledge of the substance

The second path —which in numerical terms will probably be the majority— is to operate AI without yourself having the knowledge the AI provides. Some emerging figures:

• The prompt engineer. Specialization in formulating well what's asked of a model, without necessarily understanding what the model computes.
• The doctor who delegates diagnosis. Keeps the clinical role —physical exam, patient context, therapeutic decision— but the working-up of the differential diagnosis is done in conversation with an assisted system. Hua and colleagues, in World Psychiatry 24(3), 2025, documented it in mental health; there are analogues in cardiology, radiology, dermatology.
• The lawyer who delegates the draft. The initial drafting of contracts, briefs, opinions is done by the model. The lawyer reviews, adjusts, signs. The assumption of responsibility stays with the lawyer; the elaboration of the product doesn't.
• The teacher who delegates preparation. The lesson plan, the exercises, the detailed explanations are prepared by the model. The teacher executes in the classroom and modulates by interaction.
• The journalist who delegates research. Source gathering, the first version of the article, translations, basic fact-checking are delegated. The journalist interviews, contextualizes, signs.

The operator, in all these cases, exercises a nominal function without sustaining the underlying knowledge. The system works because there's a human layer that assumes responsibility and modulates the interaction. But if the human layer is removed, the system keeps producing —probably with greater error, but it keeps going. The operator's function is as much technical as it is symbolic and legal.

This path is the one most like what Carr described in The Glass Cage (W. W. Norton, 2014) talking about airline pilots: the operator keeps responsibility and professional recognition, but their effective skill shrinks with each year in the automated cockpit. The contemporary commercial pilot flies manually a few minutes of each flight. When the unexpected happens —Air France 447, 2009— the latent capacity doesn't always respond as it should.

The operator is a viable path in the short and medium term. In the long term, when regulation accommodates and responsibility is distributed differently, part of the operator's work may disappear too. But that's a ten-to-fifteen-year scenario, not three-to-five.

Path 3: learning as identity

The third path is rarer to name in contemporary language, but it's happening. Some people learn not to be competitively more useful, but because learning is part of who they are. Reading, studying, going deep into something, not as professional investment but as a constitutive exercise.

This path isn't new. It was, largely, the premodern conception of learning. Greek paideia, German bildung, Renaissance humanist culture —all conceptions where learning was to form the subject, not to accumulate competitive advantage. Industrial modernity turned learning into pragmatic investment —«study to get a good job». The AI era, paradoxically, may return to learning part of its premodern meaning, because the pragmatic investment is being devalued.

What this implies concretely:

• Reading classic literature because it shapes you, not because it makes you competent. The emotional and cognitive investment in a nineteenth-century Russian novel doesn't pay off professionally. It pays off in identity terms —if such a thing is even measurable.
• Studying languages out of love for the language, not because you'll use them —machine translation will cover them adequately in almost all practical contexts. But learning German to read Kafka in German still gives something translation doesn't deliver.
• Making music, painting, sculpture, not to make a living from it —generative AI has collapsed the market in these areas for the average producer— but to constitute a personal practice.
• Researching obsolete topics because they interest you, not because they're in demand. Classical erudition, local history, regional geology, amateur archaeology.

This path requires conditions not everyone has. Free time, economic stability, a cultural frame that recognizes the value of non-productive activity. In societies where guaranteed income or similar structures emerge, this path can be a reasonable option for large fractions of the population. In societies where income stays tied to measured productivity, the path remains a luxury of minorities.

What the OECD and the WEF are proposing

The institutional frameworks have begun, with a lag, to react. The OECD, in Future of Education and Skills 2030 — Learning Compass (OECD Publishing, 2019, with updates from the 2024 Global Forum in Poland), proposes three transformative competencies that, according to the framework, should occupy the centre of education against instrumental competencies:

1. Creating new value —generating what AI doesn't generate well. 2. Reconciling tensions and dilemmas —operating in situations where there's no single answer. 3. Taking responsibility —assuming the consequences of decisions, even when someone else made them.

The World Economic Forum, in its Future of Jobs Report 2025 (January 2025), projects for 2030: 170 million jobs created, 92 million destroyed, a positive net balance of 78 million. AI and big data appears as the top growing skill. 39% of skills will change within the report's time horizon. 63% of employers cite the skills gap as the main barrier to transforming their business.

The figures are reasonable and, at the same time, don't solve the individual problem. The average professional who reads the report finds that «skills will change», but the report doesn't say exactly which concrete skills he, at 47 in his specific sector, will have to acquire. Converting the macro-report into an individual strategy is left to the individual —which is exactly the information asymmetry that makes the transition particularly hard for the majority.

What isn't being calculated

This is personal opinion but documentable. The three paths —specialist, operator, ritualist— aren't substitutable for one another and aren't equally probable for each person. The choice of which path to follow depends on:

• Age and available time horizon.
• Economic resources to invest in training or to accept a reduction in income.
• Psychological disposition toward change or toward stability.
• Sector and country where one operates.
• Luck —in a strong sense: being in a micro-sector AI doesn't touch by chance.

The serious question is what kind of society emerges from the aggregate distribution of choices. If the majority of the professional population ends up on path 2 —operators with no underlying knowledge— there's a question about who sustains structural knowledge in the medium term. If the majority ends up on path 1 —specialized on moving microfrontiers— there's a question about how the common cognitive infrastructure is sustained. If a significant fraction ends up on path 3 —learning as identity— there's a question about how that fraction is financed.

Nobody has done the aggregate calculation seriously. There are partial reports —WEF, OECD, Stanford HAI— but none addresses the central political question: which distribution is desirable and how to get there. The discussion stays at upskilling and reskilling as if they were technical solutions, when the problem is structural and the answer requires public debate that isn't yet happening.

An observation about the moment

The next ten years are the period in which the generation now aged 20 to 35 will make the individual decisions that will determine, in aggregate, what proportion of each path makes up the professional fabric of 2040. These decisions are being made with incomplete information, in a transitioning labour market, with no adequate institutional frameworks, with no proportional public debate. Which is normal in major technical transitions —it happened with industrialization, with suburbanization, with digitization— and which is also, in democratic terms, problematic.

Definitions

• Moving AI capability frontier: the line separating what AI systems do reasonably well from what they don't yet; it typically shifts every six to twelve months at a variable cadence by subdomain.
• Professional operator without cognitive substance: a professional who exercises a formal function whose elaboration they delegate to assisted systems, keeping legal responsibility and symbolic supervision without sustaining the underlying knowledge.
• Constitutive learning: learning whose main purpose isn't to produce measurable competence but to form the subject who learns; the dominant conception in premodern frames (paideia, bildung), residual in industrial modernity.
• Transformative competencies (OECD): three competencies —creating new value, reconciling tensions, taking responsibility— proposed as the centre of the twenty-first-century curriculum against more substitutable instrumental competencies.
• Upskilling / reskilling: practices of updating (raising the level in the same profession) and retraining (changing fields) usually proposed as the individual response to major technical transitions; uneven effectiveness depending on the origin and destination field.

References

• Newport, C. So Good They Can't Ignore You: Why Skills Trump Passion in the Quest for Work You Love. Grand Central, 2012.
• Brynjolfsson, E.; McAfee, A. The Second Machine Age. W. W. Norton, 2014.
• Carr, N. The Glass Cage: Automation and Us. W. W. Norton, 2014.
• OECD. Future of Education and Skills 2030 — Learning Compass. OECD Publishing, 2019; updates from the 2024 Global Forum in Poland.
• Susskind, D. A World Without Work. Metropolitan, 2020.
• World Economic Forum. Future of Jobs Report 2025, January 2025.
• Stanford HAI. AI Index Report 2026.

Going deeper

• Arendt, H. The Human Condition. University of Chicago Press, 1958. The distinction between labour, work and action —which underlies any serious discussion of the meaning of human work when labour is automated.
• Sennett, R. The Craftsman. Yale University Press, 2008. A frame on why the combination of manual skill, professional judgement and pride in the craft still matters for the constitution of the subject, even when measured productivity no longer requires that combination.

You might also like

• The end of the classic expert. Fairness was never the criterion
• Work as residue. The frontier moves every year
• Knowledge as flow. From stock to stream
• Learning to unlearn

Elsewhere

• Supporting Cognition With Modern Technology: Distributed Cognition Today and in an AI-Enhanced Future (PMC)
• Ethics of Artificial Intelligence and Robotics — section on employment (Stanford Encyclopedia of Philosophy)