The Removal of Cognitive Muscle

Human technological history can be read as a long sequence of attempts to separate intention from exertion. A human being first decides that something should happen. The human then has to make it happen. For most of history, these two functions were fused together inside the same biological organism. If a person wanted to cut a tree, the person had to know how to cut it and had to supply the muscular energy required to swing an axe. The axe altered the situation profoundly. It concentrated force. It improved leverage. It transformed the geometry of the interaction between body and tree. Yet it did not abolish the need for the human body to supply energy. The human still had to swing the axe. A tool made muscular work more effective, but the human remained the engine.

Machines changed this arrangement. The steam engine, the internal-combustion engine, the electric motor and the hydraulic actuator progressively removed muscular exertion from the human side of the equation. The operator could increasingly specify an action without physically supplying the energy necessary to perform it. A person driving a tractor does not provide the energy that ploughs the field. A person operating a crane does not lift the steel beam. A person riding a motorcycle does not propel the motorcycle by moving their legs against the road. The human supplies direction, judgment and control. The machine supplies power. This distinction is so familiar that it has become almost invisible. Modern civilization depends upon it. The human body is no longer the principal source of mechanical energy in most industrial activity.

Artificial intelligence introduces a structurally similar transformation into cognitive work. The analogy is not exact. It should not be made casually. Human thought is not a fuel tank, and neural computation is not simply a form of muscular energy. Intelligence is not a scalar quantity that can be poured into a machine in the way electricity is poured into a motor. Yet there is a useful economic and operational distinction between deciding what intellectual work should be done and supplying the computational capacity required to execute that work. For a long time, humans performed both functions. They selected the problem, decomposed it, searched memory, generated possibilities, calculated, wrote, compared, revised, checked and finally produced an artifact. Contemporary AI systems increasingly allow the human to retain the first function while delegating an extraordinary fraction of the second. The important change is therefore not merely that machines can now "think." It is that the marginal cost of cognitive exertion has begun to fall dramatically.

This is the deeper significance of modern generative AI. A calculator did not remove the need to know what calculation should be performed. A spreadsheet did not remove the need to formulate a financial model. A search engine did not remove the need to know what information mattered. Word processors did not remove the need to compose an argument. Earlier computational tools generally accelerated a cognitive process whose essential production still occurred in the human mind. Large language models and their descendants are different because the machine itself can now perform substantial portions of the process that converts an abstract instruction into a finished cognitive artifact. The human can say what ought to exist, and the machine can increasingly work out how to produce it.

The distinction becomes clearer if physical labor is examined at the level of control theory. An industrial machine has an actuator, a control system and an objective. The actuator supplies force. The controller determines how that force should be applied. The objective specifies the desired state. In a simple machine, these layers may be mechanically coupled. In an advanced machine, they are separated. A human operator can specify a target while the machine performs the lower-level control. Artificial intelligence introduces something analogous into knowledge work. The language model or agent becomes a general-purpose cognitive actuator. It can search, summarize, write, calculate, classify, generate code, inspect documents, reason through alternatives, operate software and increasingly invoke external tools. The human supplies the objective and constraints. The machine supplies much of the intermediate cognitive execution.

This is why the conventional metaphor of AI as merely a "tool" is inadequate. A hammer is a tool because the human must perform almost all of the relevant intelligence and the hammer supplies a very narrow physical transformation. A compiler is a more interesting tool because it performs a substantial transformation that would otherwise require human labor. A modern AI system sits further along this spectrum. It is not merely an extension of the hand. It is an extension of the process by which an intention becomes an executable intellectual object. The difference matters because the domain of possible operations is enormously larger. A hammer performs one family of physical transformations. A general-purpose model can operate across language, mathematics, software, images, audio, research and structured decision-making.

The last five years have therefore been unusual not simply because AI systems became more capable, but because the unit economics of cognitive production changed. Frontier systems in 2026 can perform many well-scoped tasks at levels comparable to or exceeding human experts. The 2026 AI Index reports that several frontier models meet or exceed human baselines on PhD-level science questions, multimodal reasoning and competition mathematics. It also reports that benchmarks are being saturated increasingly quickly, making measurement itself more difficult. (Stanford HAI) The International AI Safety Report reaches a complementary conclusion. General-purpose AI systems can now perform a wide range of well-scoped tasks with high proficiency, including coding, multilingual interaction, realistic image and video generation, and graduate-level mathematics and science. But their capabilities remain "jagged." They can fail on deceptively simple tasks, generate false statements and become unreliable in multi-step projects. (International AI Safety Report)

That qualification is essential. The claim is not that AI possesses an undifferentiated quantity of cognition that exceeds humanity in every respect. It does not. Current systems remain brittle in ways that humans are not. They can produce extraordinary solutions and then make elementary mistakes. They can solve a difficult theorem while misunderstanding a mundane instruction. They can write thousands of lines of functioning software and introduce a trivial but consequential error. They can reason impressively about a problem while hallucinating a citation that does not exist. Their competence is highly dependent on task formulation, tools, context, verification and the surrounding agentic architecture. "Raw cognitive power" is therefore a metaphor, not a scientific quantity. It refers to the amount of useful cognitive work that can be generated per unit of human attention and per unit of time.

In that narrower and more defensible sense, the transformation is profound. A human cognitive worker has a limited biological throughput. Attention fluctuates. Working memory is small. Fatigue accumulates. Reading is slow. Writing is slow. Calculation is slow. Recall is imperfect. A human can usually maintain only a limited number of independent reasoning processes at once. An AI system can produce many candidate analyses in parallel. Multiple instances can inspect different hypotheses. Agents can run searches while other processes write code. Models can repeatedly revise an artifact without experiencing boredom. They can operate at machine speed for portions of a workflow. They can be copied almost instantaneously. They can be connected to enormous stores of information and computational tools. The result is not simply a smarter assistant. It is a new production function for cognition.

The analogy with machinery becomes especially powerful here. Consider the history of transportation. Before the engine, a human who wanted to travel had to supply muscular energy directly. The horse changed the scale. The carriage improved the transmission of force. The engine then separated locomotion from the human body almost completely. A driver could operate a machine whose power was orders of magnitude greater than the driver's own muscles. The driver still needed to know where to go. The driver still needed to steer. The driver still needed to understand traffic and react to hazards. But the driver's body ceased to be the principal source of propulsion.

AI does something structurally comparable to the intellectual worker. The human remains the source of purpose. The human identifies the problem. The human decides that a particular question is worth asking. The human establishes constraints. The human decides what counts as success. But the human increasingly does not need to perform every intermediate act of cognition personally. The model can generate the draft. It can explore the alternatives. It can write the program. It can inspect the data. It can propose the experiment. It can search the literature. It can critique the first answer. It can produce a second answer. It can test the code. It can explain the result. It can repeat the cycle.

The consequence is that the scarce resource shifts. In the old cognitive economy, cognition itself was scarce. A capable analyst had to spend hours reading. A lawyer had to spend hours reviewing documents. A programmer had to spend hours writing code. A scientist had to spend hours searching the literature and constructing hypotheses. A manager had to spend hours preparing analyses and reports. The bottleneck was often the number of high-quality cognitive operations a person could perform in a working day. AI attacks precisely this bottleneck.

This is visible already in software engineering. Contemporary coding agents can perform much of the execution associated with a software task after a human establishes the objective. Anthropic's 2026 analysis of approximately 400,000 Claude Code sessions found that humans generally make most of the planning decisions while Claude makes most of the execution decisions. More domain expertise is associated with more work being accomplished per instruction. (Anthropic) This is almost a direct empirical illustration of the distinction between direction and execution. The human says, in effect, "this is what the system should accomplish." The AI increasingly decides how to implement the instruction.

The distinction also appears in measurements of agentic capability. METR's 2026 task-horizon measurements estimate how long a human expert would take to complete software tasks that frontier AI agents can successfully complete at specified reliability levels. METR explicitly warns that this is not the amount of time an AI autonomously operates in the world. It is a measure of task difficulty. Nevertheless, the trend is important. Frontier agents can now handle tasks whose human completion times extend into the multi-hour range, while successful AI execution is often substantially faster than human execution. METR also cautions that measurements at the highest horizons become noisy because the available task suite is approaching saturation. (Metr) The important fact is not a particular number. It is that the boundary between "AI can assist with this" and "AI can execute this" has moved outward.

The same phenomenon is appearing outside programming. AI systems are being incorporated into scientific workflows. Google's 2026 Co-Scientist system, for example, uses multiple AI agents to generate, debate and evolve hypotheses for scientific problems. (Google DeepMind) Systems such as AlphaEvolve have been used to discover and optimize algorithms across mathematics, computer science and infrastructure. (Google DeepMind) These systems do not eliminate the scientist. They change the scientist's position in the production process. The scientist increasingly specifies the research question and evaluates the significance of proposed solutions while AI systems conduct portions of the combinatorial search.

The phrase "combinatorial search" is important. Much of intellectual work consists of traversing a space of possibilities. A human researcher may consider ten hypotheses because considering ten thousand is impractical. A human programmer may test several approaches because testing hundreds is too expensive. A human strategist may compare five scenarios because constructing fifty detailed scenarios would take too long. AI reduces the cost of traversing such spaces. The machine can generate many candidate solutions, eliminate some, refine others and return a smaller set for human inspection. This is analogous to increasing the horsepower of a machine, but it may be more profound because cognitive search spaces are often combinatorial.

The human therefore moves upward in the abstraction hierarchy. When the machine is weak, the human specifies individual operations. When the machine becomes stronger, the human specifies intermediate objectives. When the machine becomes stronger still, the human specifies desired outcomes and constraints. This is what happened in physical machinery. An early machinist might manually control every component of a process. A modern industrial system can be given a high-level production specification and execute thousands of low-level operations automatically. The operator does not become irrelevant. The operator's job becomes more supervisory and more architectural.

This is also why prompting is a transitional concept. The prompt is the cognitive equivalent of a control signal. Early generative AI required humans to specify the desired output with considerable precision because the system had little capacity to infer intent. As models improve, the amount of detail required to produce a useful result decreases. The prompt therefore becomes less like a command to a calculator and more like a statement of objective to an increasingly autonomous worker. The important skill is no longer merely knowing which words cause a model to produce a desired response. It is knowing what objective should be given to the system in the first place.

This changes the meaning of expertise. In the pre-AI cognitive economy, expertise was partly the possession of procedural knowledge. A good lawyer knew how to conduct legal research. A good programmer knew how to implement a feature. A good analyst knew how to construct a model. A good writer knew how to organize prose. AI makes much of this procedural knowledge externally executable. It does not make the knowledge worthless. It changes its economic role. Knowing how something is done becomes less valuable when the procedure can be delegated. Knowing what should be done, why it should be done, what constraints matter and whether the result is correct becomes more valuable.

The distinction resembles the difference between a pilot and an engine. The pilot does not need to generate thrust. The engine does that. But the pilot must understand navigation, weather, aircraft behavior and the consequences of error. The greater the power of the engine, the more consequential mistakes in control can become. A weak machine may simply fail to accomplish the task. A powerful machine may accomplish the wrong task extremely efficiently.

This is the central danger of cognitive automation. The problem is not only whether AI can execute an instruction. The problem is whether the instruction describes the right objective. An AI system can be exceptionally competent at solving a badly specified problem. It can optimize the wrong metric. It can faithfully produce an elegant answer to a question that should never have been asked. It can write a perfect implementation of an incorrect business requirement. It can produce a persuasive argument from a false premise. It can search an enormous space of possibilities while moving farther and farther from the thing that actually matters.

The motorcycle analogy therefore contains an important truth but also an important limitation. A motorcycle still requires a human rider because the rider continuously perceives the environment, interprets it and controls the machine. The motorcycle supplies mechanical power but not general-purpose situational intelligence. Contemporary AI agents increasingly blur this separation. They can observe digital environments, choose actions, use tools, inspect intermediate results and revise their plans. The human does not necessarily remain in the loop at every moment. Instead, the human may establish a goal and permit the agent to operate for a period before reviewing the outcome.

That is why "human-in-the-loop" is becoming less precise as a description of advanced AI systems. The human may instead be "human-on-the-loop." The distinction matters. A human-in-the-loop system requires human authorization for each important action. A human-on-the-loop system allows the machine to execute a sequence of actions while the human supervises at a higher level. As systems become more capable, the latter arrangement becomes economically attractive. Constant human approval would destroy much of the benefit of automation. But removing approval also transfers responsibility toward objective specification, monitoring and system design.

The International AI Safety Report's 2026 assessment captures this tension. It describes AI agents as increasingly capable and reliable while emphasizing that they remain prone to basic errors that constrain their usefulness. It also identifies reliability challenges, malicious use, cyber risks and loss-of-control concerns as important emerging issues. (International AI Safety Report) The relevant conceptual point is that increasing cognitive power does not automatically produce increasing reliability. A machine can become more capable while remaining imperfectly controllable. Indeed, greater capability can make the consequences of small specification errors larger.

The economics are equally important. If AI removes the cognitive exertion required for a task, then the cost structure of that task changes. Suppose a company previously required ten analysts to perform a recurring research process. If an AI system allows two analysts to supervise the same amount of cognitive production, the organization does not simply possess a faster version of the old workforce. It has a different production function. The limiting factor may shift from labor hours to compute, data, access to proprietary information, organizational coordination or managerial judgment. Stanford's 2026 AI Index reports that productivity gains have been largest in structured and measurable work, with reported gains of 14–15% in customer support, 26% in software development and 50% in some marketing-output measures. It also finds that AI agent deployment remains relatively early even as generative AI adoption has become widespread. (Stanford HAI)

This helps explain why the labor-market effects of AI may appear strangely uneven. If AI replaces cognitive exertion rather than entire occupations, then occupations will not disappear uniformly. A job is usually a bundle of tasks. Some tasks require judgment. Some require routine execution. Some require physical presence. Some require interpersonal trust. Some require legal accountability. Some require taste. Some require institutional knowledge. AI can attack the execution component without necessarily replacing the entire occupation. The result may initially look like augmentation. But if the execution component is large enough, the number of people required to produce a given quantity of output can fall.

The first-order effect is therefore not necessarily unemployment. It is leverage. One person can do the work that previously required several people. The second-order effect is competition. If everyone receives similar cognitive leverage, the relative value of the human contribution shifts toward whatever remains scarce. The third-order effect is organizational redesign. Companies may begin constructing workflows around small groups of humans supervising large populations of AI agents. The fourth-order effect is institutional. Education, professional licensing, career ladders and apprenticeship systems built around the gradual acquisition of procedural competence may become less economically coherent.

This last point is particularly consequential. Human expertise has historically been transmitted through practice because practice was necessary to acquire the ability to perform the work. A junior programmer wrote simple programs before writing complex systems. A junior analyst built small models before building large ones. A junior lawyer reviewed simple documents before handling complicated matters. The practice itself was productive. AI changes this relationship. A novice can ask a powerful model to perform work that would previously have required years of training. This creates enormous leverage, but it also creates a paradox. If the novice delegates the work through which expertise is normally acquired, the novice may become dependent on a system whose output they cannot independently evaluate.

The 2026 AI Index explicitly identifies this possibility. It notes that productivity gains can be substantial while also pointing to evidence that heavy reliance on AI may carry long-term learning penalties. (Stanford HAI) The danger is not merely that people will become intellectually lazy. The deeper danger is that the economy may reduce the amount of human practice required to produce immediate output while simultaneously reducing the number of humans capable of independently judging that output.

This produces a new kind of scarcity: epistemic supervision. If AI can generate ten thousand plausible answers, the valuable human is not necessarily the person who can generate the eleven-thousandth answer. It may be the person who can determine which of the ten thousand answers is correct. If AI can produce a hundred business strategies, the valuable person may be the one who understands the market sufficiently to identify the strategy that will survive contact with reality. If AI can generate a scientific hypothesis, the scientist's scarce contribution may be determining which hypothesis is experimentally meaningful. If AI can write a legal argument, the lawyer's scarce contribution may be understanding the institutional and factual context that determines whether the argument is actually useful.

This is a transition from production to selection. It is also a transition from execution to judgment.

But selection should not be confused with passive approval. Human judgment itself requires models of reality. One cannot reliably evaluate an AI-generated medical analysis without medical knowledge. One cannot evaluate a software architecture without understanding software engineering. One cannot evaluate a legal argument without understanding law. The human does not become an uninformed manager standing above an intelligent machine. The human must often become a more sophisticated evaluator precisely because the machine has become more capable.

This creates a counterintuitive relationship between expertise and automation. The more powerful the machine becomes, the less the expert may need to perform routine work personally. But the more important the expert's remaining judgment may become. A master architect does not need to draw every line by hand. A master engineer does not need to perform every numerical calculation manually. Yet their understanding of structure becomes more important, not less. AI may push knowledge workers toward a similar role. The human becomes less of a producer of intermediate representations and more of an architect of the overall solution.

The architecture metaphor is useful because modern cognitive work increasingly resembles systems engineering. A human defines the objective function. The AI explores the solution space. External tools provide data and execution. Other AI systems may critique or verify the result. The human establishes constraints and handles high-consequence decisions. The final product emerges from the interaction of these components. The individual human is therefore no longer the entire cognitive stack. The human becomes one component in a larger socio-technical system.

This also changes what "intelligence" means operationally. Traditional intelligence tests mostly ask what an individual can do under controlled conditions. The emerging question is different. What can a human achieve when connected to a large-scale cognitive system? A person with mediocre coding ability may nevertheless build a functioning application with an advanced coding agent. A researcher with limited programming expertise may perform sophisticated computational analysis through natural-language interaction. A manager may generate analyses that would previously have required a specialist team. The relevant unit of capability becomes the human-machine system rather than the isolated human.

The implications resemble those of the industrial revolution, but with one crucial difference. Mechanical machines primarily amplified or replaced physical force. Cognitive machines operate on the production process that determines how other machines are used. Intelligence sits upstream of physical production. A tractor replaces muscular effort in agriculture. AI can potentially redesign the algorithm that controls the tractor, optimize its routes, analyze its crops and determine when it should operate. A machine can replace labor at one layer while AI can operate across many layers of the production hierarchy.

This is why the distinction between physical and cognitive automation may eventually break down. AI is already moving toward embodied systems. Google's 2026 Gemini Robotics work, for example, combines visual understanding, spatial reasoning and planning with systems capable of producing motor commands for robots. (Google DeepMind) The International AI Safety Report nevertheless notes that general-purpose AI systems cannot yet reliably integrate with robotic components to perform basic physical tasks such as housework. (International AI Safety Report) The direction is therefore clear without implying that general-purpose physical automation has already arrived.

If cognitive automation and physical automation converge, the historical separation between "machine as muscle" and "AI as mind" becomes less useful. A sufficiently capable system could receive an objective, reason about it, operate software, control machinery, inspect the physical environment and revise its plan. The human contribution would increasingly consist of defining objectives, constraints, values and acceptable risk. The machine would increasingly occupy the execution layer.

This brings us to the deepest question. What exactly remains human when execution becomes cheap?

The answer cannot simply be "creativity." AI systems are already creative in the ordinary operational sense. They can generate designs, stories, hypotheses, algorithms and conceptual combinations. Nor can the answer simply be "intelligence." AI systems already exhibit many forms of sophisticated reasoning. Nor is it enough to say "judgment," because judgment can itself be partially automated. The more durable distinction may be agency. Humans possess stakes in the world. Humans care about outcomes. Humans live inside institutions, bodies, relationships and histories. A machine can optimize an objective, but the question of which objectives ought to govern human life remains fundamentally political, ethical and existential.

This does not mean that AI cannot influence or even partially determine objectives. It plainly can. Recommendation systems already shape attention. Algorithms already determine prices, rankings and resource allocation. AI systems can suggest goals that humans adopt. As systems become more persuasive and more autonomous, the boundary between choosing an objective and accepting an objective generated elsewhere becomes increasingly important.

The human burden therefore becomes more difficult, not necessarily lighter. If the machine supplies cognitive force, the human must decide where that force should be applied. A powerful engine makes steering more important. A powerful weapon makes target selection more important. A powerful computer makes problem selection more important. A powerful AI makes objective selection more important.

This is the central asymmetry of the new cognitive economy. Execution becomes abundant. Direction becomes scarce.

But even "direction" is too narrow. The human must specify the destination, define the constraints, recognize when the destination is wrong, evaluate trade-offs, decide when evidence is sufficient, determine which risks are acceptable and remain accountable for consequences. The human must know enough to detect when the machine is confidently wrong. The human must understand the environment in which the output will operate. In many high-stakes domains, verification becomes inseparable from expertise.

This is why the future of knowledge work may not be characterized accurately as humans versus AI. The more useful distinction is between humans who can direct cognitive machinery effectively and humans who cannot. The competitive unit may become a person or organization with a high-quality objective, excellent domain knowledge and access to powerful agents. Such a person may possess enormous cognitive leverage. A person without the ability to formulate objectives or evaluate outputs may possess access to the same machinery while gaining much less from it.

The consequence is a peculiar reversal of the old industrial hierarchy. During industrialization, physical strength became less important because machines supplied it. Cognitive supervision became more important because somebody had to operate and coordinate the machines. AI may now do something analogous to cognition itself. The person who previously supplied the intellectual energy becomes the operator of an external cognitive engine.

The metaphor of the motorcycle therefore captures something essential. The rider does not make the motorcycle move. The engine does. The rider decides where to go, interprets the road, manages risk and prevents catastrophe. But the analogy must be extended for AI. An advanced AI agent is not merely a motorcycle. It is closer to a motorcycle whose engine can also read the map, propose destinations, inspect the road, choose routes, diagnose mechanical problems and occasionally decide that another route would better satisfy the stated objective. The human remains responsible for the journey, but the machine increasingly participates in determining how the journey is conducted.

That is why the distinction between direction and execution will become one of the defining distinctions of the AI era. The human advantage will increasingly lie not in the ability to produce every intermediate step, but in the ability to specify the right end state. The machine advantage will increasingly lie in traversing the enormous space between the initial instruction and that end state.

There is, however, a final complication. Humans themselves are not static. Once AI becomes part of the cognitive environment, human cognition will adapt around it. The boundary of competence will move. People will learn to formulate problems differently because they know machines can solve certain classes of problems. Organizations will change their workflows. New professions will emerge around evaluation, orchestration, auditing, context engineering, system design and agent supervision. Some existing professions will contract. Others will become more productive. The resulting equilibrium cannot be inferred simply by asking which occupations AI can perform. The more important question is what happens to the price and availability of the cognitive services those occupations provide.

The industrial worker did not disappear when the engine arrived. The worker became part of a different production system. The engineer, operator, mechanic, manager and designer became more important in some respects and less important in others. Entire categories of work emerged that would have been unintelligible before industrial machinery. The same process is likely to occur with cognitive machinery.

The historical analogy also warns against a simplistic assumption that automation necessarily eliminates all human effort. When a task becomes cheaper, demand often expands. When computation becomes cheap, people perform more computation. When communication becomes cheap, people communicate more. When software development becomes cheaper, organizations may build more software rather than simply employing fewer programmers. AI could therefore increase the total volume of cognitive production even while reducing the amount of human labor required per unit of output. Stanford's 2026 data already show widespread adoption alongside still-early agent deployment, suggesting that the economy remains in a transition rather than a settled post-labor state. (Stanford HAI)

The ultimate transformation may therefore be less about the disappearance of work than about the disappearance of cognitive exertion as the principal constraint on work. For centuries, the human mind was both the architect and the workshop. The person had to conceive the solution and labor through its intermediate steps. AI begins to separate these functions. The mind can specify the objective. The machine can become the workshop.

This is the real conceptual discontinuity.

A hammer amplified the hand. An engine replaced the muscle. A computer amplified calculation. The internet amplified access to information. Modern AI begins to externalize the process by which information is transformed into reasoning, language, code and plans. It is therefore not simply another instrument in the historical inventory of tools. It is a machine for producing intermediate cognitive states.

That distinction explains both the extraordinary promise and the unusual danger. If the bottleneck is no longer the amount of cognitive work available, then humanity can attempt projects previously limited by intellectual throughput. Scientific search can expand. Software production can expand. Engineering design can expand. Administrative processes can expand. Small teams can accomplish tasks that once required large organizations. The productive potential is enormous.

But the bottleneck moves upstream. The scarce resource becomes judgment about what should be produced. If humanity can summon vast quantities of cognitive execution almost instantly, then the ability to identify worthwhile objectives becomes disproportionately important. The problem is no longer merely that humans cannot think fast enough. It may become that humans cannot decide what is worth thinking about.

The deepest lesson of the machinery analogy is therefore not that humans will become obsolete. It is that humans should expect their economic role to migrate toward the layer that machines have not yet made cheap. In physical automation, that layer was control, maintenance, design and ownership. In cognitive automation, it is increasingly objective formation, contextual judgment, verification, accountability and the selection of ends.

The old worker had to possess both the blueprint and the muscle. Industrial machinery removed the muscle from the human body. AI is beginning to remove the cognitive exertion from the human mind. The blueprint remains, but even that metaphor is becoming unstable because increasingly capable systems can help generate blueprints themselves.

What remains unmistakably human is not the ability to perform every step. It is the responsibility to decide which steps ought to be performed, toward which ends, under which constraints, and for whose benefit.

That is the new frontier of labor.

The machine has taken the cognitive muscle.

The human still has to choose where to swing.