2.6.1 Demand Expansion

First, consider demand expansion. Automation in a sector increases productivity, which lowers prices and can expand demand for the sector’s output (Autor Reference Autor2022). In some cases, demand increases can actually outpace supply increases–the “Jevons effect” (Alcott Reference Alcott2005). In the present case, cheaper cognition might expand what’s worth cognizing.

This mechanism is real. But on the assumption that AI capabilities keep improving, demand expansion likely can’t keep pace. Here’s why: As AI gets better at cognitive tasks, productivity gains compound. For demand expansion to prevent displacement, demand would need to grow at least as fast. But there’s no reason to think demand for legal analysis or medical diagnosis expands without limit simply because it gets cheaper. Instead, at some point, we’ve presumably done all the legal analysis worth doing. Put in economic terms: demand expansion absorbs productivity gains only if price elasticity of demand persistently exceeds one. But demand curves for cognitive services aren’t infinitely elastic! And at compounding rates, they don’t need to be steep for long before reaching saturation.

2.6.2 Task Complementarity

Second, consider task complementarity. Automation of some tasks can increase the value of complementary tasks (Milgrom and Roberts Reference Milgrom and Roberts1990). Everyone’s favorite example is ATMs: When ATMs automated cash handling, they made bank branches cheaper to operate. Banks responded by opening more branches and hiring more workers to handle loan consultations and complex account issues. Similarly, if AI handles a subset of cognitive tasks, human judgment on the remaining cognitive tasks could become more valuable.

There are two versions of this idea. The weaker version holds that AI capabilities will plateau at routine cognitive tasks, while complex judgment remains durably human. This would create a stable refuge. But AI that improves on existing trends year over year is unlikely to remain confined to routine cognitive work.

The stronger version says that even as AI capabilities expand, the remaining cognitive tasks assigned to humans become more valuable precisely because they’re complements to an ever-growing AI-performed portion. This is correct as far as it goes: Complemen-tarity can technically raise the marginal product of the remaining human contributions indefinitely even under continuation. But indefinitely growing marginal product per worker is perfectly compatible with sharply reduced demand for workers overall. Indeed, it is what we would expect. This possibility is not theoretical. It is the story of American manufacturing: output up, productivity up, and employment down by two-thirds. Indefinite complementarity might therefore create a smaller corps of highly valuable human cognitive workers, but this wouldn’t preserve cognitive work at scale. As I’ll argue below regarding O-ring effects, mechanisms that make remaining workers more productive in general can't explain why existing employment levels in cognitive work would persist.

2.6.3 Wealth Effects

Third, consider wealth effects. Automation generates wealth through productivity gains, and wealth generates demand for goods and services, including labor-intensive ones (Autor Reference Autor2022). If AI dramatically increases productivity in cognitive work, the resulting wealth could fund employment in sectors far removed from the automated tasks themselves.

I expect economists are correct that a world where all cognitive work is automated is a much – maybe much much – richer world. But setting aside distributional questions, wealth effects are perfectly compatible with the collapse of cognitive work as a distinctive category. For instance, AI-driven productivity gains might create wealth that funds more massage therapists and home health aides. But that would not be the preservation of cognitive work as we know it! It would be the widespread displacement of human beings from cognitive work into other kinds of work. If anything, wealth effects facilitate displacement by ensuring displaced cognitive workers can find other employment – just not in cognitive work.

2.6.4 Comparative Advantage

Fourth, consider comparative advantage. Comparative advantage says that what matters for task allocation isn’t who performs a task best in absolute terms but who is relatively better (cf. Buchanan Reference Buchanan1991). Intuitively, if a senior lawyer is 3× faster than a junior at drafting but only 2× faster at client meetings, it is relatively costly for the senior to spend time in meetings (since they could be drafting instead). Substitute AI for the senior lawyer, and this might suggest humans might retain whatever cognitive tasks AI is relatively less dominant at.

Comparative advantage only operates under constraints imposed by scarcity. What would make the capacity of AI systems to accomplish cognitive tasks scarce? A single AI system can’t do task A and task B simultaneously. But this trivial bottleneck only helps humans if scaling or multitasking is constrained. To make the argument for displacement as difficult as possible, suppose there are persistent architectural constraints on multitasking.

That leaves constraints on firms’ ability to deploy more than one AI system. What could impose such constraints? The most obvious candidates are scarcity in both compute and electricity. But these constraints are rapidly loosening, not tightening: data centers are being built at unprecedented scale, power infrastructure is expanding, and efficiency gains in algorithms mean that each unit of compute accomplishes more.Footnote ⁴

But suppose scaling does face some binding constraints in the form of relatively scarce compute and electricity. Even so, it seems to me that the alleged refuge offered by comparative advantage to human cognitive workers will be cold comfort. There are two reasons for this.

First, for comparative advantage to preserve human roles in cognitive work, humans’ relative disadvantage must be smallest in at least some cognitive tasks. But why think this is likely to be true? These systems are being deliberately designed to efficiently accomplish cognitive tasks in particular. If anything, a very different outcome seems much more likely, given the capabilities trajectory we’re assuming here: humans retain comparative advantage at various not-cognitive tasks (robotics is hard) even as their disadvantage compared to AI systems continues to grow for cognitive tasks in particular. Worse, even if humans retain comparative advantage with respect to some cognitive tasks, why think this much smaller subset of cognitive tasks will be large enough to support the status quo for cognitive work?

But suppose (what seems incredible) that human cognitive workers retain comparative advantage with respect to most cognitive tasks. Still, this tells us nothing specific about the wage those workers will demand in an unrestricted market for their labor. That wage will be capped by the opportunity cost of deploying an AI system instead. But as AI gets increasingly cheaper and increasingly abundant, that opportunity cost drops. Correspondingly, the ceiling on human cognitive workers’ wages will predictably drop, potentially far below subsistence (Korinek and Suh Reference Korinek and Suh2024).

2.6.5 O-Ring Automation

Finally, consider a recent proposal involving concentration effects (Gans and Goldfarb Reference Gans and Goldfarb2026). Gans and Goldfarb show that when production requires multiple tasks whose qualities multiply – an “O-ring” production function, following Kremer (Reference Kremer1993) – automating some tasks can raise the marginal product of remaining human labor. Intuitively: if a worker allocates fixed time across tasks, automating only some tasks lets the worker concentrate remaining time on fewer tasks, raising quality on each. Since output is multiplicative in task qualities, this concentration can scale the value of human work in remaining bottleneck tasks. But given our assumption that AI capabilities keep improving, this mechanism fails to preserve the status quo for cognitive work for two reasons.

First, the O-ring mechanism operates at the level of individual workers, not labor markets. The model therefore explains why a certain kind of concentration in production functions makes each employed worker more valuable. But it tells us nothing about how many such workers an economy needs. As we noted earlier, indefinitely growing marginal product per worker is compatible with sharply reduced demand for workers overall.

Second, Gans and Goldfarb’s model assumes that task quality equals labor productivity times hours, with no diminishing returns (Gans and Goldfarb Reference Gans and Goldfarb2026, 3). This is a bizarre assumption if we are trying to model cognitive work in particular. Their model works if we relax this assumption to allow for concave quality functions (i.e., we allow diminishing returns in quality for time invested). But this doesn’t resolve the underlying problem, which is that, intuitively, there’s only so much extra human attention you can give a cognitive task before you hit a hard quality ceiling.

More carefully: Gans and Goldfarb’s model shows that when workers are displaced from some (cognitive) tasks, the remaining tasks are “performed with disproportionately high care and are, therefore, disproportionately hard to replace” (Gans and Goldfarb Reference Gans and Goldfarb2026, 10).Footnote ⁵ Suppose that’s true. Still, most cognitive tasks intuitively face very steep diminishing returns or outright quality ceilings. The existence of these ceilings matters for the automation “barrier” that Gans and Goldfarb identify. The lower the quality ceiling, the lower the barrier to automation.

So, how low is the quality ceiling on cognitive tasks humans currently perform? Put otherwise: suppose a median- or extremely high-skill cognitive worker had infinite time to devote to some moderately or extremely demanding cognitive task. How much higher quality should we expect the output to be? Maybe twice for the median, and somewhat higher at the extreme, depending on the kind of task. Surely not more than one order of magnitude at any widely distributed skill level. If this is right, then an O-ring barrier to automation is a correspondingly easy barrier to clear, at least for the bulk of cognitive tasks.

2.6.6 Sources of Friction

The analysis so far assumes labor markets operate without major institutional interference and that standard economic dynamics largely control the outcome: firms substitute AI for human cognitive labor (almost) whenever doing so is cheaper (controlling for the opportunity cost of deploying the AI elsewhere). But labor markets don’t operate in a vacuum. Various forms of friction might slow or redirect displacement even under continuation. Three kinds of friction are especially salient.

First, governments could restrict AI’s role in cognitive work in various ways (e.g., via licensing requirements). This is the most immediate political lever available for stalling the displacement of human workers from cognitive work (Acemoglu and Restrepo Reference Acemoglu and Restrepo2020a). But regulatory intervention requires sustained political will against prevailing economic headwinds. Individual jurisdictions face competitive pressure compounded by automated cognitive work’s lack of geographic constraints. Restrictive nations risk firms relocating cognitive tasks to lighter-touch jurisdictions, much as manufacturing was off-shored (Blinder Reference Blinder2006). In the absence of coordinated international action, regulation fragments and likely slows rather than outright prevents displacement.

Legal frameworks potentially create a second source of friction. Existing legal systems assign responsibility to corporations and to human beings. When AI systems make errors, liability can be uncertain: Does responsibility fall on the developer, the deployer, or the nominal human supervisor (Matthias Reference Matthias2004)? This uncertainty may make firms cautious about automation. But legal frameworks standardly adapt to economic pressure. In particular, new liability regimes emerge and insurance markets can be expected to develop in ways that permit automation to proceed (Trout 2025). Moreover, as Bayern (Reference Bayern2016) has shown, it has been technically possible for decades to turn any instance of AI into a legally responsible entity without changing a single law.

Third, various groups of workers might organize collective resistance. But collective action faces significant defection problems. Individual firms or professionals who adopt AI in cognitive work are likely to gain competitive advantage over those who resist. Some sectors with outsized cultural or political influence may hold out longer. For an example of each, SAG-AFTRA’s recent strike provisions restricting AI use in production and longshoremen’s decades-long resistance to port automation show that organized labor can slow adoption in specific sectors. But organized labor seems unlikely to be successful at halting the widespread economic transformation of work this time around, just as it was unsuccessful in halting previous technological transformations of work.

Indeed, all these dynamics have operated during earlier technological transitions. Craft guilds resisted mechanization in the nineteenth century. Regulatory frameworks struggled to accommodate factory automation. Liability questions slowed adoption of new industrial technologies: Railroad accidents in the nineteenth century drove the development of modern liability insurance. But in each case friction simply stretched the transition without stopping it. The same pattern seems likely here.

2.8.1 What Is Cognitive Assistance?

In 1997, Garry Kasparov lost at chess to IBM’s Deep Blue. Thereafter, Kasparov proposed formalizing a new format: “advanced chess,” where human players could consult computers during matches. In advanced chess, human–computer teams (“centaurs” or “cyborgs”) played better chess than either humans or computers alone. They divided the cognitive labor of chess: each side did what it was better at. Centaur chess is a paradigmatic case of technological assistance in the performance of a complex, multifactor, highly context-sensitive (though typically very low-stakes) cognitive task.Footnote ⁶

The centaur (“augmentation”) model has therefore become the default frame for thinking about AI assistance in cognitive work (Acemoglu and Restrepo Reference Acemoglu and Restrepo2020a; Autor Reference Autor2024; Lei and Kim Reference Lei and Kim2024). The vision is appealing: AI handles the tedious parts of cognitive work while humans contribute judgment, creativity, and oversight. But “AI handles the tedious parts” is too vague to be analytically useful. Which parts, exactly, can we expect to be offloaded, and which are likely to remain human?

The answer turns on what makes cognitive work difficult in the first place: what we can think of as humans’ cognitive bottleneck. As the case of chess nicely illustrates, humans face hard constraints on both the scale and the intensity of their effortful cognition. For example, we can hold only a few (dozen) items (e.g., states of the board) in working memory at once (Baddeley and Hitch 1974). Intensity, sustained attention (e.g., to a game) degrades over time (Monsell and Driver Reference Monsell and Driver2000), and switching between tasks (e.g., between move recall and move evaluation) incurs cognitive costs (Uncapher et al. Reference Uncapher, Thieu and Wagner2016) that require recovery time to fully refresh.

More generally – that is, beyond just chess – effortful cognition requires exercising a cluster of related meta-cognitive capacities. This suite of capacities is what cognitive scientists call our “executive system,” and it comprises “executive functions” that can only do one thing at a time and typically not for that long (though these constraints are liable to training) (Diamond Reference Diamond2013; Ahmed et al. Reference Ahmed, Tang, Waters and Davis-Kean2019). For instance, capacities such as working memory, sustained attention, inhibitory control, and planning are among those that manage and coordinate conscious thought; these capacities all operate serially rather than in parallel (Miyake et al. Reference Miyake, Friedman, Emerson, Witzki, How-erter and Wager2000). Moreover, a well-functioning executive system is itself quite cognitively demanding to acquire and requires explicit, deliberate training (as compared to the implicit, informal training needed to develop cognitive capacities like language) (Center on the Developing Child 2011; Miyake et al. Reference Miyake, Friedman, Emerson, Witzki, How-erter and Wager2000).

Cognitive work is cognitively demanding precisely because it demands serial processing through this single-channel human cognitive bottleneck. For example, legal reasoning, programming, medical diagnosis, and strategic planning each require holding multiple considerations in mind, resisting multiple distractions, searching systematically through large information spaces, and monitoring each part of a long chain of reasoning for error. Even the most capable, best-trained humans can only do so much of this at once, and only for so long. (I return to cognitive difficulty and executive functions in Section 4.)

What, then, is artificial cognitive assistance? I propose the following functional definition: artificial cognitive assistance is assistance that aims to route around the human cognitive bottleneck in some way.

Given this definition, whether something counts as cognitive assistance depends on context. This variability is a feature of the definition: assistance is relational, indexed to the cognitive constraints of workers (e.g., particular workers on particular tasks).

Such assistance can take many different forms, depending on the details of the cognitive work in question, including the broader sector of work in which it is embedded. This is why, increasingly, companies are building bespoke cognitive assistants for different domains (Sharadin Reference Sharadin2026b; Gabriel et al. Reference Gabriel, Manzini and Keeling2024; Eloundou et al. Reference Eloundou, Manning, Mishkin and Rock2023).

Return to centaur chess. The bottleneck framing explains why the human–computer division of labor takes the form it does. The goal is high-quality chess. The bottleneck is human cognition. The centaur configuration is how you route around it while keeping humans in the loop. Computers handle exhaustive search tasks that would overwhelm human limits. Humans contribute strategic judgment. Chess engines therefore count as cognitive assistants in our sense.

Now forget centaur chess. The same structure appears wherever AI assistance is being deployed in cognitive work: centaur everything.

In radiology, AI processes full scans and flags regions of concern–handling the systematic visual search that would otherwise load working memory and sustained attention. This frees radiologists to concentrate on evaluating the flagged candidates (Boeken et al. Reference Boeken, Feydy and Lecler2023). In legal research, AI searches exhaustively through case law while tracking relevance criteria, freeing lawyers to focus on evaluating which precedents actually matter for the case at hand. In software development, AI generates candidate code, handling the working-memory load of syntax and edge cases, freeing engineers to concentrate on architectural judgment and verification.

The subject matter varies, but the structure doesn’t: AI takes on parts of tasks that would otherwise tax executive function, freeing humans to concentrate their limited cognitive resources on what remains.

2.8.2 Why Will Cognitive Assistance Be Ubiquitous?

Cognitive assistance in cognitive work is assistance that aims to route around human cognitive bottlenecks in some way – the human–centaur X configuration, for whatever X is relevant. Why think this configuration will be increasingly commonplace? Why expect ubiquitous cognitive assistance?

Two forces, technological defaults and economic incentives, drive ubiquity, and together they create a ratchet.

Start with defaults. Defaults will structure people’s choices about how (and when) to use this cognitively assistive technology in exactly the same way they systematically shape our choices in other areas.Footnote ⁷ Assistance will be the default. We know this because in domains where AI assistants are most mature, assistance is already part of the default interface for cognitive work. For instance, productivity suites have persistent “Help me X” for some more or less general cognitive X (e.g., write, generate ideas, draw, code) buttons. In some cases, users cannot adjust the default. This is in fact how the tools most workers use to complete cognitive tasks already now work, by default.

This emerging default status is not an accident; it is by deliberate design. Google’s researchers are especially explicit about the future they’re building; in a recent paper, they describe AI assistants as “the main interface of the future” (Gabriel et al. Reference Gabriel, Manzini and Keeling2024, p. 15). The business models of dominant technology companies demand that AI assistance becomes the standard interface for cognitive work.

Moreover, defaults, as we know, can easily harden into norms (Bovens Reference Bovens, Grüne-Yanoff and Ove2009; Thaler and Sunstein Reference Thaler and Sunstein2021). The social norm these companies seem to envision is one where unassisted cognitive work marks someone as an eccentric, like refusing to use a calculator, insisting on paper maps, or ignoring social media.

Economic incentives, meanwhile, accelerate these dynamics. Early adopters (driven by defaults) gain an edge; competitors adopt to keep pace; eventually nonuse marks a worker as incompetent. This is an arms race with a predictable outcome.

In general, we can expect corporate actors to shape a path of least (in this case, cognitive) resistance and set it as the default; economic incentives will then ensure that people walk it. Some individuals will resist on principle. Some domains will remain stubbornly resistant to assistance. But the forces are strong enough to push most cognitive workers toward ubiquitous use. And as I’ll now explain, ubiquitous use is sufficient to transform cognitive work.

2.8.3 How Will Ubiquitous Cognitive Assistance Affect Cognitive Work?

The most obvious impact of ubiquitous cognitive assistance on cognitive work is that it will disrupt our fifth feature: limited productivity growth. Centaur–Magnus can play many more simultaneous games at a higher level of quality than plain old human Magnus.

This is good news! Baumol’s cost disease is a disease: nobody thinks it’s a good thing that, for example, education and healthcare continue to grow relatively more expensive over time. Curing the disease means productivity can finally grow in sectors where it has historically stagnated.

Does this destabilize the other features defining cognitive work? Let’s work through each.

High skill requirements will likely persist in an ubiquitously assisted world, since assisted work still requires expertise. Lawyers reviewing AI output must know law. Maybe more intuitively, I want my centaur-radiologist to be, you know, a well-trained, highly skilled radiologist! The specific skills will likely shift – more verification, less generation – but cognitive work is unlikely to become recognizably “low” skill. If it did, why wouldn’t the already quite capable AI be doing it?

What about credentials? This feature likely faces significant disruption. Credentials currently strongly signal a capacity for, among other things, generation. For example, law school exams test brief-writing, medical boards test diagnosis, and programming interviews test code production. But assisted work is more likely to emphasize verification: evaluating AI output rather than producing it from scratch. For some period, credentials will therefore signal the wrong (at least one much less important) thing: skills correlated with what the assistant now handles rather than skills required for working effectively with assistance. Eventually, filtering and credentialing institutions such as universities will presumably adapt (cf. Arrow Reference Arrow1973).

In an ubiquitously assisted world, automation resistance is, of course, significantly weakened. But assistance might stabilize at a human-in-the-loop equilibrium. If, for example, verification genuinely requires human judgment that AI cannot replicate, the centaur configuration could persist rather than sliding toward full displacement.

As I think of things, the crucial question is what happens with wage premiums. Productivity gains mean – potentially much – more output per worker. If demand for cognitive output doesn’t expand proportionally (I’ll say more about this below), then significantly fewer workers will be needed to meet demand. Significantly fewer workers needed to meet demand, same supply of trained workers, standard labor market dynamics: significant downward pressure on wages.

So the question isn’t whether ubiquitous assistance would transform cognitive work: it would, by enabling significantly increased productivity growth in sectors that have suffered Baumol’s cost disease for decades. The question is whether that transformation cascades into the package’s other features – especially wage premiums – or whether it gets absorbed by countervailing forces. The next section considers reasons to doubt both the ubiquity of assistance and the severity of the cascade.

2.9.1 Denying Ubiquity

One might argue that capable assistive systems will exist but won’t be ubiquitous, reducing both their positive impact on productivity and their negative impact on wages. I don’t have much to say about this possibility.

Some cognitive work might escape assistance entirely. Perhaps there are some economically valuable tasks with no relevant human cognitive bottleneck to route around. I expect there are not many such tasks that currently exist (though see below on new tasks). Maybe tasks requiring judgments of taste are like this. But such tasks are typically embedded in workflows full of bottleneck-heavy cognitive tasks. Certainly, some institutions and sectors will resist adopting cognitive assistance for reasons beyond competitive pressure, and they will pay a premium (e.g., in growth or profits) for it. But if assistance becomes ubiquitous in most cognitive work while artisanal cognizers hold out, this will not preserve cognitive work as we know it.

2.9.2 Denying the Effect on Work

You might instead accept that assistance will be ubiquitous but resist the claim that it will have the two effects I’ve claimed it will on cognitive work. There are two salient possibilities.

The first involves imagining worlds in which, although assistance is ubiquitous, it simply has no positive impact on productivity. For reasons I’ve already explained, this seems unlikely: These systems are already capable of performing a range of economically valuable cognitive tasks to a very high degree of reliability. Admittedly, this has not yet showed up in measures of per-worker productivity. But this is likely best explained by diffusion dynamics and well-known problems with measuring changes to per-worker productivity caused by changes in technology (Brynjolfsson Reference Brynjolfsson1993).

The second possibility is more interesting: productivity gains occur, but wages don’t collapse. Several dynamics might produce this outcome.

First, consider the possibility of new tasks emerging. This is a standard economist’s reply to pessimism about automation: technological change historically creates new tasks even as it automates old ones (Acemoglu and Restrepo Reference Acemoglu and Restrepo2019). But the relevant question here isn’t whether productivity gains from automation will lead to the creation of new tasks, including tasks that require human labor. Suppose new tasks emerge broadly and quickly enough that unemployment stays below 10%. The relevant question here is whether these new tasks will be cognitive tasks that can be bundled into jobs that thereby preserve the distinctive package associated with cognitive work. Here I’m skeptical.

The mechanism that generates new tasks – roughly, productivity gains creating surplus, surplus funding new activities – doesn’t preferentially create high-skill, credential-gated, wage-premium work, let alone work with those features that involves sitting in front of a keyboard. It creates whatever work the economy demands. If that’s care work, personal services, entertainment, or some other new difficult-to-imagine kind of not-cognitive work, then new tasks will absorb the surplus of labor generated by ubiquitous assistance without preserving cognitive work as an economically relevant category.

I myself think we should be agnostic about the skill content, credential requirements, and wages of tasks that don’t yet exist. By contrast, someone interested in resisting the effects of assistance on cognitive work who predicts that new tasks will emerge to absorb productivity gains needs to explain why cognitive tasks in particular will be the ones that so emerge. Crucially, the burden of proof is then on this person to do something that economic history teaches us it’s extraordinarily difficult to do: they must predict the specific skill content of new tasks (Autor et al. Reference Autor, Levy and Murnane2003).

Now consider the possibility of demand expansion in a world of ubiquitous assistance. As we saw earlier, in worlds of displacement, expanded demand for cognitive output can simply be met with more AI. But in a world of ubiquitous assistance, expanded demand will by hypothesis require more human-AI “centaurs” – which, of course, means more humans. For example, suppose cheaper legal analysis leads to vastly more demand for legal analysis that requires human verification. Then, productivity gains might translate into more assisted lawyers rather than fewer well-paid ones. Whether this dynamic effectively preserves cognitive work depends on two things. First, assistance must remain stable rather than sliding toward displacement. Second, demand elasticity must remain high enough that expanded output absorbs the productivity gains. Both of these possibilities are very uncertain, but neither is crazy. I lack the space to consider them in detail here.

Finally, consider the effect on cognitive work of ubiquitous assistance in noncompetitive sectors. Large portions of cognitive work occur in sectors insulated from competitive pressure to minimize labor costs. For instance, education, healthcare, government, and nonprofits employ substantial fractions of cognitive workers. But wages in these sectors are typically set by, for example, politics, collective bargaining, and institutional norms rather than marginal productivity calculations. Nevertheless, it seems unlikely that these sectors can preserve cognitive work’s wage premiums indefinitely if the broader cognitive labor market collapses.

3.2.1 The Empirical Argument

What makes work good for people, assuming it sometimes is? Here, we can safely ignore the effect of money, since there’s obviously nothing about cognitive work per se that guarantees better pay than not-cognitive work. The usual (nonpecuniary) candidates for the goods of work are things such as engagement, achievement, social connection, autonomy, and purpose (cf. Crawford Reference Crawford2009; Danaher Reference Danaher2017; Gheaus and Herzog Reference Gheaus and Herzog2016; Cholbi Reference Cholbi, Edward and Nodelman2023). If cognitive work had special access to these goods, that could explain why the displacement of humans from cognitive work is a distinctive threat to human flourishing. But it doesn’t. Each of these goods appears in not-cognitive work.

Consider engagement first. Csikszentmihalyi’s research on flow states found deep engagement as readily in carpentry as in coding (Csikszentmihalyi and Csikzentmihaly Reference Csikszentmihalyi and Csikzentmihaly1990; Csikszentmihalyi et al. Reference Csikszentmihalyi, Abuhamdeh and Nakamura2014). The absorption that makes work satisfying doesn’t track the cognitive content of tasks. Instead, it tracks the match between challenge and skill (Csikszentmihalyi Reference Csikzentimihalyi1975).

The goods associated with achievement follow the same pattern. A master craftsman achieves as genuinely as a master analyst (Sennett Reference Sennett2008). For instance, an electrician might troubleshoot a baffling fault by exercising judgment. This counts as achievement on even the most demanding accounts (Bradford Reference Bradford2015; Hurka Reference Hurka2020).

Social connection forms around shared challenges and mutual reliance, neither of which requires specifically cognitive tasks. Workplace bonds on a construction site are at least as deep as those in consultancies.

Autonomy is often cited as the most important good (other than money!) associated with work (cf. Gheaus and Herzog Reference Gheaus and Herzog2016). But cognitive work has no monopoly on autonomy. Instead, worker autonomy varies more by job structure than by the cognitive content of job-related tasks. A tenured academic has autonomy; a junior associate billing hours at a partner’s direction doesn’t. Journeyman plumbers have it; call center workers don’t. Autonomy at work intuitively tracks hierarchy and independence, not the cognitive content of job-related tasks.

What about purpose? I doubt this is a construct for which we have a good empirical measure: purpose is woefully under-specified as an empirical construct (cf. Rosso et al. Reference Rosso, Dekas and Wrzesniewski2010). (I return to this point in Section 5.) But intuitively, purpose, if anything, favors not-cognitive work. It would be strange to claim that plumbing or nursing involves less purpose than consulting or copywriting.

The goods that make work good (when it is) appear across both cognitive and not-cognitive work. (Notably, so do the bads: alienation and tedium afflict cognitive and not-cognitive workers alike.)

3.2.2 The Conceptual Argument

Perhaps the empirical argument misses something. Maybe there’s a conceptual route to the conclusion that humanity’s displacement from cognitive work would be a special loss.

If cognitive work is uniquely valuable for human flourishing, then something beyond its economically distinctive features must make it so. What could that be? A natural idea: cognitive work involves thinking, and thinking is arguably distinctively human. So perhaps a loss of cognitive work is a loss of an activity that’s especially human. Let’s grant that thinking is in some way especially human. Still, this doesn’t explain why a loss of cognitive work is distinctively bad.

This is because not-cognitive work involves a lot of thinking, too. A carpenter deciding how to approach a tricky joint is engaged in planning; a mechanic reasoning backward from symptoms to causes is engaged in problem-solving. The difference between cognitive and not-cognitive work tends to lie in the object of thought, not in the presence or absence of thought per se.

At this point, a defender of cognitive work’s special status might try a different move. Perhaps what makes not-cognitive jobs good, when they are good, is precisely their cognitive elements. The carpenter’s job is good for them, or promotes their living a flourishing life, because of the planning and problem-solving it involves – that is, the cognitive parts of the job. The “not-cognitive” label that we apply to these jobs is in this respect misleading; the especially valuable core of these jobs is cognitive after all.

This has a whiff of circularity. It amounts to saying: a specific kind of cognitive activity is what makes work valuable; so, whenever we have the intuition that allegedly not-cognitive work is valuable, that must be because of some of that work’s (latent) cognitive components. But this begs the question. Why think planning is what makes carpentry valuable (pleasurable, contribute to flourishing, and so on) rather than its physicality? The move also carries an unpleasant odor of elitism – as if the philosophers and analysts get to claim the good parts of manual labor as secretly their own, while the sawing and hammering are mere dross.

3.2.3 From Existence to Parity

The arguments so far establish that the goods typically associated with work exist in not-cognitive work. But, one might object, existence is strictly weaker than parity. There are at least two relevant possibilities.

First, perhaps these goods are systematically less accessible in not-cognitive work. Maybe flow states are rarer on construction sites than in design studios, or the conditions for autonomous judgment are harder to achieve when work is physical. This is an empirical question, and as we’ve seen, the evidence doesn’t support the asymmetry (Dormann and Zapf Reference Dormann and Zapf2001; Humphrey et al. Reference Humphrey, Nahrgang and Morgeson2007).

Second, and more interestingly, one might argue that different kinds of work have internal goods specific to the practice (MacIntyre Reference MacIntyre2007). On this view, what makes legal reasoning valuable for a lawyer isn’t generic “achievement” but something intrinsic to the practice of law – a form of excellence that can’t transfer to plumbing, however excellent the plumbing. Displacement from one practice to another might therefore involve genuine loss even if both practices offer engagement, autonomy, and the rest.

This objection has real force, but it cuts against special treatment of cognitive work. If internal goods are practice specific, then displacement from any practice to any other involves losing access to the relevant practice-dependent goods. A displaced carpenter loses the internal goods of carpentry just as a displaced lawyer loses the internal goods of law. This is a reason to care about displacement from specific practices generally, not a reason to think displacement from cognitive work is distinctively threatening. Moreover, the internal goods of a practice are goods of the practice, not straightforwardly goods for the practitioner. The excellence internal to law matters to law; whether it matters to the lawyer’s welfare is a further question that the generic goods are better positioned to answer.

3.2.4 The Theoretical Argument

We can confirm the parity claim by examining major theories of well-being directly. Hedonists care about pleasure, which, as we’ve seen, tracks challenge–skill match rather than cognitive content (Gregory Reference Gregory and Fletcher2015; Dormann and Zapf Reference Dormann and Zapf2001). Preference theorists defer to what people actually want, and preferences about work are diverse – where preferences for cognitive work exist, they typically turn out to be instrumental, tracking wages and working conditions rather than cognitive activity as such (Heathwood Reference 80Heathwood and Fletcher2015; Deci and Ryan Reference Deci and Ryan2000). Objective-list theorists point to goods such as achievement, knowledge, and autonomy (Fletcher Reference Fletcher and Fletcher2015) – goods the empirical argument has already shown appear in both categories. No major theory of well-being identifies cognitive work as distinctively good for people.

3.2.5 Summing Up the Case for Parity

Each route leads to the same place. The goods associated with work appear empirically in both cognitive and not-cognitive work (Table 3). There’s nothing conceptually special about “cognitive” that could plausibly ground an evaluative distinction between cognitive and not-cognitive work. And no major theory of well-being identifies cognitive work as intrinsically better for people. The parity claim holds: cognitive and not-cognitive work are on par with respect to the (noneconomic) goods of work. For this reason, the widespread displacement of human beings from cognitive work is not, as such, a loss.

3.3.1 Research and Discovery

Section 1 identified limited productivity growth as a distinctive feature of cognitive work. As we’ve seen, Baumol’s cost disease makes cognitive-labor-intensive goods steadily more expensive because productivity can’t keep pace with wage growth (Baumol Reference Baumol1967, Reference Baumol2012).

Cost disease isn’t the only constraint on productivity growth afflicting sectors that rely on cognitive work. Research productivity per worker has been declining for decades: The number of researchers required to achieve a given rate of productivity growth has increased dramatically over time (Bloom et al. Reference Bloom, Jones, Reenen and Webb2020; Gordon Reference Gordon2016; Kortum Reference Kortum1997). Ideas, as Bloom et al. put it, are getting harder to find.

These two constraints compound each other. Cost disease means that the cognitive labor required for research is expensive and trending more so. Declining research productivity means we need even more of that expensive labor to achieve the same results.

The widespread displacement of humans from cognitive work attacks both constraints, though in different ways.

Replacing human cognitive labor with AI attacks cost disease directly by massively expanding the effective supply of cognitive labor.

Displacement may also attack declining per-worker research productivity. The fact that ideas are getting harder to find arguably reflects the increasing difficulty of navigating complex search spaces. As fields mature, the unexplored territory that might contain breakthroughs becomes harder to identify and reach. Human researchers, with finite careers and cognitive limitations, can only explore so much (Bloom et al. Reference Bloom, Jones, Reenen and Webb2020).

AI systems may be differentially well suited to this kind of exploration. Consider drug discovery. Traditional pharmaceutical research involves scientists generating hypotheses based on existing knowledge and intuition, then testing candidates through slow (and expensive) experimental pipelines (DiMasi et al. Reference DiMasi, Grabowski and Hansen2016). Systems such as diffusion models for molecular design do something different: They explore chemical space in ways humans couldn’t, identifying candidate molecules that human researchers would never have considered (Watson et al. Reference Watson, Juergens and Bennett2023).

These systems don’t themselves displace researchers – they don’t use themselves! But if displacement proceeds as Section 2 suggests and capabilities continue to improve, the systems doing the exploring will increasingly be the ones deciding what to explore by themselves using other AI tools.

Google’s “AI co-scientist” (Gottweis et al. Reference Gottweis, Weng and Daryin2025) offers a glimpse of what this future might look like. Consider the example from Google’s paper introducing the system: A team of researchers recently spent several years piecing together the solution to a biological puzzle involving bacterial parasites and arrived at their breakthrough only through what they themselves described as serendipity. Before their results were published, Google gave the same puzzle to co-scientist, which they had designed to mirror the kind of cognitive labor scientists do when developing hypotheses about novel biological phenomena. The system’s top-ranked hypothesis was identical to the human team’s breakthrough (Penadés et al. Reference Penadés, Gottweis and He2025). Importantly, this is a first-generation system: To my knowledge, it’s the first attempt by anyone to build a system designed to displace human beings from these kinds of complex cognitive tasks.

Breaking Baumol’s cost disease gives us more cognitive labor at the frontier; there will be more “minds” working on problems, without the training pipeline that currently limits supply and without the associated costs of employing human researchers. This is CEO of Anthropic Dario Amodei’s “country of geniuses in a datacenter” (Amodei Reference Amodei2024). But if the frontier is genuinely hard, if good ideas require searching vast spaces that human cognition isn’t well suited to navigate, then more human-style cognition will also hit diminishing returns. We may need not just more cognition but different kinds of cognition. AI that explores differently, not just faster, could steepen the curve of discovery itself (Cockburn et al. Reference Cockburn, Henderson and Stern2018).

The welfare implications of this are significant. Research generates compounding returns: breakthroughs enable further breakthroughs (Jones Reference Jones2005; Romer Reference Romer1990). And these benefits persist through time: a new class of antibiotics saves lives for decades. Accelerating research is about as high leverage an intervention for improving long-run human welfare as it’s possible to imagine.

3.3.2 Access to Justice

There’s a separate reason to welcome displacement: in many societies, completing certain cognitive tasks is a prerequisite for fair institutional treatment, and automating those tasks would democratize access to justice.

Legal services are the obvious case. If legal tasks are expensive, justice is correspondingly expensive to access. The poor and otherwise vulnerable therefore predictably get less access to justice (Rhode Reference Rhode2004; Sandefur Reference Sandefur2019). Displacement from legal labor therefore matters because fair institutional treatment shouldn’t depend on ability to pay.

The kinds of cognitive tasks involved in tax preparation and benefits administration illustrate the same point. The American tax code – like many (but not all) tax codes in developed economies – is baroque (Hoppe et al. Reference Hoppe, Schanz, Sturm and Sureth-Sloane2023). Completing the cognitive tasks required to conform to it in the most efficient way demands either significant expertise and the time to deploy it or the money to hire an expert’s time. The same is true of applications for various kinds of government programs: They tax applicants’ time and discourage eligible people from claiming what they’re entitled to (Lowrey Reference Lowrey2021; Herd and Moynihan Reference Herd and Moynihan2019). And in these cases, the largest burdens of compliance fall hardest on those with the fewest resources to shoulder them. Intuitively, wealthy individuals have relatively complex taxes but are eligible for fewer benefit programs with high administrative burdens; by contrast, poorer individuals have relatively simple taxes but are eligible for relatively more high-burden benefit programs.

We could address these problems by eliminating the relevant administrative “sludge” (Sunstein Reference Sunstein2021, Reference Sunstein2022). But that requires a very specific kind of political action. We could also subsidize human help, for example, by expanding legal aid and funding “benefits navigators” or providing free tax preparation services. But human cognitive labor is expensive and scarce.

Assistance could perhaps ease these constraints. But even dramatically more productive pro bono lawyers can only serve so many clients, and even dramatically cheaper lawyers are still too expensive for many who need help. Systems capable of displacing human cognitive workers, meanwhile, change the economics qualitatively. When AI performs these tasks without human cognitive labor, the marginal cost of helping one more person shoulder some unfair administrative burden is compute and electricity.

3.4.1 Not Enough Work

The most commonly cited worry about displacement of humans from cognitive labor is that there won’t be anywhere for displaced cognitive workers to go. Previous technological transitions had destinations: agricultural workers moved into manufacturing, and manufacturing workers moved into services and cognitive work. But cognitive work was supposed by many to be the final destination – the refuge for workers displaced from everything else (Mokyr Reference Mokyr2011). If it’s automated too, where do people go?

There are two versions of this worry.

The first is that there simply isn’t enough not-cognitive work to absorb displaced cognitive workers. This is very uncertain, though I’m inclined to be optimistic. Not-cognitive work as we’ve defined it here includes the skilled trades, personal services, entertainment, physical labor in unstructured environments, large portions of healthcare, and much else. Whether this sector can absorb a large influx of former cognitive workers depends on elasticity of demand, the pace of transition, and policy choices about work-sharing and labor markets (Acemoglu and Restrepo Reference Acemoglu and Restrepo2020a; Autor Reference Autor2024).

There are hard questions about how to make the transition to not-cognitive work easier for individual cognitive workers. But I doubt we should be concerned that there won’t be “enough” work to go around. The “lump of labor” fallacy is supposed to be a fallacy precisely because new work typically emerges as old work is automated (Autor Reference Autor2015). Whether this pattern continues under the AI-driven displacement of cognitive work is uncertain. But if it does, then the concern is about managing the transition smoothly, not about displacement being bad as such. Faster transitions mean more acute disruption, but the disruption is the same kind previous generations faced. And we have better tools now than, for example, nineteenth-century agricultural workers had; in particular, we have safety nets and the resources to fund them (Esping-Andersen Reference Esping-Andersen1990). The question is whether we deploy these tools adequately, which is a political question, not a reason to think displacement from cognitive work into not-cognitive work is uniquely threatening.

A second version of this worry is that future AI systems will eventually displace not-cognitive work too, foreclosing lateral moves by workers entirely. This version of the worry is significantly more speculative. In Section 2 I noted that embodied tasks in unstructured environments remain harder to automate than those that involve pure cognition. But set that aside. Suppose embodied AI systems (e.g., humanoid robots) eventually displace humans from not-cognitive work as well. Then the concern is that humans have nowhere left to sell their labor. I’ll say two things about this more speculative possibility.

First, it seems obvious to me that the mere fact that no entity is economically incentivized to purchase any human’s labor is not itself a special moral reason for concern. Instead, this tends to be morally concerning when, for instance, this fact attends various problematic distributions of resources. I address such worries about distributive justice below.

Second, I’m tempted to go further. The argument for the parity claim aimed to show that not-cognitive work offers the same menu of goods arguably required for human flourishing as cognitive work offers. But it also implied something stronger, namely that work (in the economic sense) isn’t strictly necessary for human flourishing. After all, there’s no reason to think that being paid is a necessary condition for enjoying any of the (noneconomic) goods that we noted are typically associated with (both cognitive and not-cognitive) work (things such as achievement, social connection, and so on).

In any case, notice that in this second version of the worry, the concern is no longer about displacement from cognitive work specifically, but about displacement from work as such. That concern turns on a different question – whether performing remunerated tasks is necessary for human flourishing. I am extremely skeptical, but addressing this issue in any detail would go beyond the scope of the present discussion.

3.4.2 Distributive Justice

Someone might admit that displacement could be good but only if the benefits associated with displacing human cognitive workers from cognitive work are shared. Unfortunately (the thought goes), they won’t be. Capital will predictably capture all (or most) of these benefits; meanwhile, displaced human workers (and their families) will be immiserated. I’ll make two initial comments designed to allay the concern, and then explain why I think this is not best understood as a concern about displacement per se.

First, the natural distribution of displacement’s benefits deserves careful attention. One of these benefits is potentially very meaningful productivity gains in sectors currently suffering from Baumol’s cost disease. Take one such sector: medical research. Productivity gains in medical research yield better medicine. And the benefits of better medicine don’t flow primarily to capital; they flow to patients. For example, a new cancer treatment primarily benefits patients who have cancer, not shareholders, even if the treatment is sold on the open market for a profit. This remains true even if the opportunity to access the treatment is deeply inegalitarian: Welfare gains from medical research aren’t captured by capital in the way profits are. Similar remarks apply to other sectors afflicted by low productivity. The sludge-clearing benefits noted above are even more clearly egalitarian: The whole point is that administrative burdens fall hardest on those with the fewest resources, so the immediate effect of removing them is to help the worst-off most.

The distributive worry makes most sense for productivity gains that show up as corporate profits. It makes much less sense for gains that show up as, for example, cures, or as the removal of barriers to fair treatment.

Still, the concern isn’t baseless. Some of the gains from automating human cognitive labor will almost inevitably take the form of profits, and the distribution of those profits is arguably (Nozick Reference Nozick1974) a legitimate concern. But, and this is my second comment, this represents a familiar problem of distributive justice. In fact, it’s almost exactly the same problem of distributive justice that arose with industrialization, with globalization, and with every previous wave of automation. We have myriad frameworks for thinking about what a fair distribution of the surplus generated by cooperation requires (Rawls Reference Rawls1971; Arneson Reference Arneson1989; Parfit Reference Parfit2018; Frankfurt Reference Frankfurt2018) and many known mechanisms for achieving it (James et al. Reference James, Retting, Shogren, Watson and Wills2022; Van Parijs and Vanderborght Reference 87Van Parijs and Vanderborght2017). The question is whether we use them.

The structure of this objection is revealing. It concedes that the widespread displacement of human beings from cognitive work plus a fair distribution of the benefits thereof yields a good outcome. It then predicts that for some reason we won’t achieve such a fair distribution. But that’s a (pessimistic) political prediction, not a claim about displacement itself. If the prediction is right, the correct response is to work harder on distributive politics, not to prefer a world where displacement doesn’t happen and the relevant gains don’t exist to be distributed. In general, the correct response to a bad distribution of some good is a better distribution of that good.

So, at least as I understand things, there’s nothing about displacement causing unjust distributions that makes those distributions a morally distinctive problem for displacement itself. Consider an analogy. Suppose a company invents a drug that makes humans effectively immortal. Left to natural market forces, this would presumably have very seriously inegalitarian effects on the already very unequal global distribution of wealth. But it would be very misleading to put this point by saying things like: there’s a serious moral problem with making humanity immortal. Equally, in the case of widespread displacement.

3.4.3 Autonomy and Distribution

Suppose some distribution is adequate in material terms and meets whatever further evaluative standard (if any) we think an economic distribution must meet. Still, something troubling could remain. Under wage labor, people’s material welfare depends on their ability to sell something others need. This arrangement has many problems. But it arguably gives people a kind of independence. If displacement extends beyond cognitive work to labor generally – or if the not-cognitive sector can’t absorb displaced cognitive workers – then people’s welfare will depend on the continued willingness of whoever controls productive capacity to share it. Perhaps there’s something problematic about this kind of dependence. I am skeptical, for three reasons.

First, the autonomy of wage labor is easy to romanticize. The freedom to sell your labor provides little independence if you’ll starve without a buyer; in that case, unrestrained buyers will dictate terms. And buyers – that is, employers – are in fact largely unrestrained (cf. Anderson Reference Anderson2017). This is why most workers have a right of exit from their employment in name only. Dependence on labor markets may therefore be no better than dependence on distributive institutions.

Second, the worry here is really about institutional design, not (even total) displacement as such. Concentrated ownership of AI plus weak institutions for governance could indeed leave people more vulnerable than they are now. But that’s an argument for better institutional design and better governance, not an argument against displacement itself. There’s no shortage of proposals for improving governance along whatever dimension of freedom we think matters; these face political obstacles, but they’re not conceptually mysterious.

Third, dependence as such is arguably not worth caring about. For example, republican political philosophers such as Philip Pettit argue that what makes dependence objectionable is arbitrariness (Pettit Reference Pettit1997). The relevant question on this view isn’t whether we depend on others (or on institutions) but whether that dependence is structured by accountable institutions or left to the discretion of the powerful (Pettit Reference Pettit2012).

3.4.4 Loss of Control

Kulveit et al. (Reference Kulveit, Douglas, Ammann, Turan, Krueger and Duvenaud2025) argue that even incremental AI displacement could permanently erode humanity’s influence over systems we depend on. Let’s grant that this result would be bad. Is this a reason to think total displacement of humans from cognitive work would be bad? No. Let me explain.

Societal systems currently align with human interests partly through explicit channels (voting, consumer choice) but more fundamentally through their implicit dependence on human participation. For instance, states need citizens to generate taxable income, and markets need workers to produce and consumers to buy. This dependence creates alignment between human preferences and social systems not by explicit design but because systems that fail to support human flourishing lose the human participation they require. Widespread displacement threatens to sever this link. If AI generates most economic value, the historical connection between taxation and representation weakens. If AI handles most labor, markets optimize for AI activity rather than human preferences. The systems keep functioning – but they no longer need us, and so they no longer need to serve us: we are, in Kulveit et al.’s evocative phrase, “like cattle in an industrial farm – fed and housed by systems they neither comprehend nor influence” (Kulveit et al. Reference Kulveit, Douglas, Ammann, Turan, Krueger and Duvenaud2025, 5).

Being treated the way cattle are treated in industrial food production is, again, an admittedly very worrying possibility! (Notably, consensus on this point is some reason to think we are failing in our moral duties to cattle raised for slaughter.) But notice what is required to arrive at this outcome: economic displacement broad enough that human participation becomes completely dispensable for economic production in roughly the same way the cattle’s preferences are treated as completely dispensable in terms of our decisions about how to treat them. (I set aside the cultural dimensions Kulveit et al. discuss, since economic displacement does the argumentative work.) This is an extremely speculative possibility. In particular, it is significantly more speculative possibility than the possibility I’m focused on here in this Element, which is (more or less) widespread displacement of human beings from cognitive work in particular. Cognitive work is roughly half of employment in developed economies, and the other half promises to be substantially harder to automate. Displaced cognitive workers who move into not-cognitive work still pay taxes and still spend their wages on goods they want. Hence I am relatively sanguine about the likelihood of Kulveit et. al’s extremely grim outcomes in the case of humanity’s widespread (even total) displacement from only cognitive work in particular. Again, things may look different in outcomes where all economically valuable work is automated.

If a residual worry involving a loss of control remains, I suspect it concerns accountability rather than the distribution of human-versus-artificial cognition as such. If AI systems controlled by a narrow elite making consequential judgments without democratic oversight would be troubling, then human elites doing the same should be equally troubling. My own view is that both these outcomes are less troubling than many think, for broadly meritocratic-cum-perfectionist reasons (cf. Brennan Reference Brennan2016; Bai Reference Bai2019). But set that to one side. The controlling question, if we are worried about this somewhat different issue, is whether crucial decision-making (human or artificial) is subject to whatever level of accountability we think is in fact appropriate – and if it isn’t, how we can ensure that, going forward, it will be. That’s a question about collective governance on which reasonable people disagree – not a reason to think displacement is per se problematic.

4.3.1 Pleasurable Difficulty

Cognitive difficulty is typically aversive (Székely and Michael Reference Székely and Michael2021). But under the right conditions – particularly when the activity demanding the cognition is freely chosen (Wahlers et al. Reference Wahlers, Garrison and Hart2025) – it can also be enjoyable. Rawls articulated a general version of this idea as the “Aristotelian Principle”: “other things equal, human beings enjoy the exercise of their realized capacities (their innate or trained abilities), and this enjoyment increases the more the capacity is realized, or the greater its complexity” (Rawls Reference Rawls1971, 373). The principle applies to capacities of all kinds, including our cognitive capacities. We often enjoy the effortful engagement of our (innate or trained) cognitive capacities.

This enjoyment appears in cognitive work when conditions permit. The key empirically robust condition is a balance between challenge and skill (Csikszentmihalyi and Csikzentmihaly Reference Csikszentmihalyi and Csikzentmihaly1990). When challenge is too low, boredom results; when too high, anxiety. The sweet spot, where difficulty stretches but doesn’t exceed capacity, produces the absorbed enjoyment that Aristotle noticed, Rawls codified, and Csikszentmihalyi popularized.

How could ubiquitous cognitive assistance threaten this value? Easy: assistance could systematically shift much cognitive work below the relevant difficulty threshold for absorbed enjoyment. As a result, cognitive work could become systematically more boring.

4.3.2 Difficult Achievement

Cognitive difficulty might also be good for a more theoretical reason. Gwen Bradford has argued that “genuine” achievement requires two things: competent causation and difficulty (Bradford Reference Bradford2015). On Bradford’s view, you achieve some outcome just in case you skillfully control the causal process that produces it and it was difficult for you to do so.

A lawyer who researches, analyzes, and drafts a complex brief has achieved something in Bradford’s sense. An assisted lawyer who prompts an AI system, reviews its output, and approves the filing has certainly done something. But has she achieved anything? Maybe not. If difficulty is constitutive of achievement, then reducing the difficulty of some task reduces the degree to which completing that task is an achievement. At the limit, if cognitive assistance eliminates the cognitive difficulty associated with a task entirely, then completing that task may not count as an achievement at all.

Why care whether assisted cognitive workers “genuinely” or “Capital A” (cf. Bradford Reference Bradford2015) achieve (as much)? Achievement appears on nearly every plausible objective list of welfare goods (Fletcher Reference Fletcher and Fletcher2015; Parfit Reference Parfit1984). If less difficulty means fewer achievements, then ceteris paribus assistance makes assisted lives go worse by depriving people of a widely recognized component of human flourishing.

There are obvious ways to resist this worry. One is to reject Bradford’s view – to deny that difficulty is constitutive of achievement. I’m sympathetic to this move on theoretical grounds: It lets us say the ubiquitously assisted achieve just as much as the unassisted, only that they do so effortlessly.Footnote ⁹ But this doesn’t dissolve the worry about value; it merely relocates it. Even if effortless achievement counts as genuine achievement, it may not contribute to welfare in the same way. For theorists who think achievement is a welfare good, what matters isn’t the achievement itself but the effort behind it. Simon Keller is explicit on this point, saying that a “plausible rough-and-ready” claim relating achievement to welfare is that “the greater the effort required for an individual to achieve her goal, the more her welfare is enhanced by its achievement” (Keller Reference Keller2004, 34).

On this view, then, reducing the effort required for achievement ipso facto reduces its contribution to welfare (whether or not it reduces the extent to which it is an achievement). Since ubiquitous assistance will by hypothesis reduce the effort required for a wide array of (more or less cognitive) achievements, it will thereby reduce the contribution these achievements make to individuals’ well-being.

4.5.1 Extended Mind

Maybe the framing of cognitive assistance I’ve offered is confused. Cognitive processes aren’t confined to the skull; instead, our cognition extends into our tools and environments. On this view, AI assistance is simply the latest extension of our minds. There’s therefore no “offloading” (Risko and Gilbert Reference 85Risko and Gilbert2016) to worry about, no substitution of external for internal, just cognition distributed differently across a hybrid human–AI system (Clark Reference Clark2010). Steve Jobs famously described computers as “bicycles for the mind” (Agrawal et al. Reference Agrawal, Gans and Goldfarb2025). Maybe worrying about cognitive ease brought on by ubiquitous cognitive assistance is like worrying that your legs are doing less work because you’re riding a bicycle. I have three responses to this line of thought.

First, what AI assistants do looks more like substitution than extension. The paradigm cases of extended cognition involve tools that augment capacities the agent is actively deploying: in Chalmers and Clark’s (Reference Chalmers and Clark1998) famous example, a notebook augments the memory of an Alzheimer patient, Otto. In this example, Otto is actively using his memory (albeit in diminished form) to store and retrieve items from the notebook. AI assistance, meanwhile, often simply replaces the agent’s cognitive engagement entirely. This is especially true when it’s being used not by adult professionals to enhance their productivity in high-skill cognitive work but by children to avoid the aversive feelings associated with cognitive difficulty (Burns et al. Reference Burns, Winthrop, Luther, Venetis and Karim2026).

Crucially, this is not an accident: substitution rather than extension is the stated aim of the largest companies developing cognitively assistive technology. For example, researchers at Google explicitly disavow the extended mind framing, saying that “we are suggesting that AI assistants are not merely external cognitive subsystems that constitute part of the user’s extended mind” (Gabriel et al. Reference Gabriel, Manzini and Keeling2024); later, they double down on this characterization, saying that the AI assistants Google aims to develop and deploy to as many individuals as possible are not properly analogized to “the notebook in the Clark and Chalmers […] example of an Alzheimer’s patient who uses a notebook as a functional substitute for biological memory” (Gabriel et al. Reference Gabriel, Manzini and Keeling2024, 15). I suggest we take the designers of these tools at their word.

Second, even if we accept the idea that this technology “extends” rather than substitutes for our own cognition, this doesn’t resolve the concerns I’ve identified. The nature of our unextended cognitive capacities can still matter a great deal! An extended cognitive system with a well-functioning human component differs in significant ways from the same system with a degraded human component. This is much like how a cyclist with strong legs differs from one with atrophied legs, even if both get around equally well in some environments. Equally, if the human substrate of executive function simply works worse in the future, then even if a perfectly ubiquitously cognitively assisted person may not experience this difference, they will still be significantly cognitively enfeebled when unassisted. We may care intrinsically about this fact, as many moral philosophers no doubt do (cf. Kant Reference Kant1785). Or we may care instrumentally about it, since on occasions where assistance is less than perfectly ubiquitous, this may cause significant harm.

Third, and crucially, when it comes to children, it is, as we’ve seen, morally illicit for adults to do things that close down opportunities for children to freely make the choice to “extend” themselves. To freely make this choice, children must be in possession of the fundamental cognitive capacities to make it. But the present trajectory is toward structuring the developmental environment in ways that make it less rather than more likely that children will in fact develop these capacities, and so eventually be in a position to freely choose for themselves whether they want to be “extended.” In my view, this is a bit like Clark and Chalmers giving poor Otto early-onset Alzheimer’s, buying him a notebook, and calling it a wash.Footnote ¹⁵

4.5.2 Difficulty Is Often Bad

One might object that I’ve romanticized cognitive difficulty. Much cognitive work is tedious and alienating. Why mourn its passing?

It’s correct that not all difficulty is valuable. Stressful cognitive overload doesn’t build executive function. The kind of difficulty that develops our cognitive capacities is appropriately challenging engagement. As I noted earlier, this can be understood broadly as exposure to cognitive tasks that stretch our capacities but don’t overwhelm them. Tedium and burnout fail this condition from opposite directions.

But ubiquitous cognitive assistance doesn’t selectively remove “bad” cognitive difficulty while preserving “good” cognitive difficulty. Capable AI assistants that can perform cognitive tasks more or less chirpily handle whatever cognitive tasks they can handle. This increasingly includes the challenging, judgment-requiring parts of cognitive tasks, which are precisely the parts that constitute appropriately demanding engagement. For adults, this may mean what remains is the residue AI can’t yet touch: for instance, a final sign-off that carries legal liability. But this isn’t a part of cognitive work that produces flow or constitutes (welfare-enhancing, “genuine”) achievement. Of course, as we’ve noted, adults can presumably find cognitive engagement elsewhere; so maybe this isn’t so prudentially bad, on net.

For children, I’ve argued the moral risk is more serious: the existing balance of opportunity and incentive for productive cognitive struggle is poised to shift dramatically. The existing equilibrium is one where children are regularly presented with many situations that demand cognitive struggle and where they are incentivized to struggle cognitively in those situations in part because it’s difficult to avoid a satisfactory result without doing so. These are the cases of (developmentally) “good” difficulty.

But what does an ubiquitously assisted future look like for children? It’s one where by definition fewer situations in children’s cognitive environments meet these two conditions – there are fewer demands and weaker incentives. Crucially, this is true even if no one optimizes for this environment.

Consider, again, the analogy to our physical environments. Cars, elevators, and suburban infrastructure weren’t designed to make children sedentary. But they helped make childhood in developed economies substantially more sedentary nonetheless. A child in 1926 walked to school, climbed stairs, and played in the street; a child in 2026 is driven everywhere, takes escalators, and mostly plays indoors.Footnote ¹⁶ The physical environment became easier to navigate without anyone intending to reduce the opportunity or incentive for children to engage in physical activity.

Instead, the reduction in opportunity and incentive for exerting physical effort was a side effect of infrastructure built for adult convenience. The same applies here. Cognitive assistants are built to make cognitive tasks easier for adults in the same way escalators are built to make stairs easier for adults. But a system designed to help lawyers draft briefs can obviously help a twelve-year-old write an essay. To imagine otherwise requires imagining either that children somehow won’t have access to the defining technology of their era (we’ll ban children from escalators) or that they’ll systematically choose the cognitively more demanding path when an easier one is available and unstigmatized (they’ll freely choose stairs). As I’ll now explain, neither idea is plausible.

4.5.3 Restricting Access to Children

Maybe you think the developmental concerns are serious but tractable through straightforward policy: restrict children’s access to enfeebling cognitive assistance until their executive functions are sufficiently developed. If ubiquitous assistance threatens childhood development, then we should treat it like we treat alcohol or tobacco (ideally also: smartphones and social media) and impose age restrictions, require parental controls, and enforce robust educational guidelines.

I have four comments on this idea.

First, implementing policy in this area requires settling on a principled age range. The developmental science points somewhere uncomfortable. The prefrontal cortex continues developing into the mid-twenties (Fuster Reference Fuster2002). Capacities such as impulse control, long-term planning, and cognitive flexibility are still being refined when people are graduating college, entering the workforce, and starting families (Smith and Jonides Reference Smith and Jonides1999). A restriction that actually tracked executive function maturity would need to extend well beyond what we legally or culturally consider childhood.

Of course, this is an old challenge that we’ve managed to solve for other developmentally harmful technology. So suppose we agree on an age. Still (my second comment), enforcement of an age restriction in this case is extremely challenging. One natural thought is that schools are the lever: ban cognitive assistance in educational settings and the problem is largely solved. But this overestimates how much of children’s cognitive development happens in school. Children spend only roughly 15% of their waking hours in formal education (Walberg Reference Walberg1984); the rest is homework, play, socializing, navigating the world, and–increasingly–interacting with digital infrastructure that contains ubiquitous cognitive assistance by default. If executive functions develop through repeated encounters with cognitive demands across contexts, then protecting the classroom while leaving the rest of childhood cognitively frictionless accomplishes relatively little.

The broader point is that AI assistance doesn’t arrive as a discrete product that can be age-gated at point of sale like tobacco or corralled onto dedicated platforms like social media. It’s embedded in digital infrastructure. Meaningfully restricting access would require coordinated action across institutions on a scale unheard of in the history of technology governance. For the record, I am largely in favor of attempting this. But it’s a big ask!

Third, even if technically enforceable, the political economy includes a bevy of perverse incentives. Suppose that we don’t get global coordination on restrictions for children’s use of these systems. Then the question is how various actors who could in principle impose meaningful restrictions are incentivized to introduce such restrictions.

Consider first the incentives facing schools and parents. Schools adopt AI tools because parents demand them, because governments encourage “future-ready” curricula (cf. Trump Reference Trump2025), and because the tools appear to improve measurable outcomes such as test scores and assignment completion rates (Burns et al. Reference Burns, Winthrop, Luther, Venetis and Karim2026). Whether these gains reflect genuine learning won’t be clear for years. Education research is notoriously slow, and controlled studies that might reveal long-term developmental effects lag far behind adoption curves. In the meantime, a school that restricts AI while competitors embrace it risks losing students to districts that seem more forward-looking. Parents, meanwhile, face their own competitive pressures: withholding assistance from your child when peers have access looks like handicapping them for college admissions and career opportunities. And these incentives for parents to demand more AI in schools feed back into school policy, either directly or indirectly (e.g., through politics).

What about governments? Most municipalities lack the capacity to set enforceable restrictions on any kind of technological diffusion; worryingly, some nations lack this capacity too. Among those capable of restricting access to new technology, competitive pressures push in the opposite direction. A country that limits AI in education while rivals embrace it risks falling behind in the global contest for economic and technological leadership. Restriction looks like unilateral technological disarmament. Of course, it’s only childhood disarmament – sounds good to me! But perhaps unfortunately, given the globally competitive landscape, it seems unlikely that most governments will agree to coordinate on this kind of restriction. This pessimistic conclusion is compatible with some governments, following Australia’s leading moral position on a symmetrical set of issues regarding social media (Nogrady Reference 84Nogrady2024; UNICEF Australia n.d.), acting unilaterally to help their own country’s children. It is also compatible with multi-lateral agreements aimed at reducing catastrophic risks from advanced AI systems.

Fourth, this idea – ban AI assistance for children – isn’t really an objection to the claims I’ve argued for here in this section: it instead concedes the core concern. If keeping cognitive assistance away from children is a good idea, then the developmental risks identified in this section are serious enough to warrant significant formal or informal restrictions on children’s access to a technology currently being built and deployed at enormous scale and in basically unrestricted fashion. To repeat: I am myself in favor of such restrictions! But as I currently understand the policy landscape, this would be a very significant shift from current policy and practice. Current policy and practice point toward increasing rather than decreasing the exposure of children to cognitive assistance.

4.5.4 People Will Choose Difficult Things

In an assisted world, people can seek cognitive difficulty voluntarily, either in cognitive work itself or outside such work. Sports, hobbies, puzzles, and games provide cognitive challenge without economic pressure. Those who value the goods provided by cognitive difficulty can pursue these goods deliberately; those who don’t can enjoy the ease (cf. Danaher Reference Danaher2017, Reference Danaher2019). What’s wrong with letting people choose? This is an optimistic picture.

For adults, there are two problems with this optimistic picture. First, ubiquitous cognitive assistance at work doesn’t magically free up time for demanding pursuits elsewhere. A (full-time) job that’s been cognitively hollowed out is still a (full-time) job. Cognitive demands may shrink, but we’ve been given no reason to think the hours will also shrink, and economic history suggests they won’t (Autor Reference Autor2015). (Unless, of course, total work hours shrink to zero. But that is a different concern.) Cognitive workers won’t be liberated to spend their newfound mental energy on challenging hobbies; they’ll simply be left with less cognitively demanding work that fills the same hours in an assisted world as it did in an unassisted one. Second, we can expect that in a cognitively assisted world choosing cognitive difficulty within cognitive work will be very strongly penalized.

For children, the optimistic picture flatly misrepresents the moral situation. While it could possibly be reasonable to blame an adult for choosing cognitive ease over cognitive effort in some contexts, this is unreasonable when it comes to children. Instead, the amount of cognitive effort it makes sense to expect children to exert is largely a function of the cognitive demands they face in their environments. These environments, in turn, are shaped by the choices of adults – parents and, to a much lesser extent, schools.

4.5.5 Executive Function Won’t Matter in Assisted Worlds

If AI handles the cognitively demanding parts of life, perhaps skill at executive function is simply less valuable than we thought. There are two versions of this objection, depending on what kind of value is at issue.

One idea is that executive function is less instrumentally valuable in a less cognitively demanding world. People arguably don’t need navigation skills in a world with GPS; maybe people don’t need planning and organization in a world with ubiquitous cognitive assistance.

To this I say: That depends on how ubiquitous the assistance is and which other features of the world are held fixed. If we imagine a world of extreme abundance – such as Pixar’s WALL-E, maybe – then perhaps the cognitive demands on individuals are truly minimal, and executive function is instrumentally worthless. If we imagine (what is more likely) a world where although work is transformed, there’s still “moderate scarcity” in material terms (cf. Rawls Reference Rawls2001), then I expect there will still sometimes be significant instrumental returns to having robust executive function. Here’s a simple argument: even the most capable, most ubiquitous cognitive assistant can’t stop a teenager from doing something rash. A well-functioning executive system can.

A second, different idea is that executive function is less intrinsically valuable in a less cognitively demanding world. But this seems confused. Among those who think the capacities involved in executive function are intrinsically good to possess, this is usually said to be because those capacities (in whole or in part) constitute a capacity for autonomy, which is itself supposed to be intrinsically good (cf. Schapiro Reference Schapiro1999; Roth and Surprenant Reference Roth and Surprenant2012). But a capacity for executive function wouldn’t cease to be a constituent part of a capacity for autonomy simply because a machine could in principle do all of our hard thinking for us. So, on the view that executive function is intrinsically good for people, it’s difficult to see how it could be less so in an ubiquitously assisted world.