No age was ever rescued by the powers it was taught to shrink. (Contra)
A curious tax argument has entered Swedish politics. Arash Gilan, entrepreneur and author of 404: Människan saknas, recently argued that Sweden should stop taxing human time and begin taxing productivity. The argument received considerably more attention after Gilan reported on LinkedIn that Sweden's centre-right vice prime minister, Ebba Busch, had discussed his book on TV4 and said that we are taxing the wrong thing. The underlying proposition is superficially compelling: if an algorithm performs in seconds what previously required hundreds of human working hours, a tax system built heavily around wages may eventually lose part of its base.
The first half of that argument deserves to be taken seriously. Sweden really does tax labor heavily, and the Swedish government's fiscal material shows how central taxes on labor, including income taxes and employer contributions, remain to public revenue. If machine execution eventually substitutes for a large fraction of human labor and national income correspondingly shifts from wages toward profits and capital, then the tax mix will eventually have to adapt. That is not a radical proposition; it is a fairly ordinary consequence of a changing tax base.
The problem begins with the proposed instrument. In economics and public policy, an instrument is simply the mechanism chosen to produce a desired effect. A diagnosis can be correct while the instrument chosen to address it is destructive. Gilan's diagnosis is plausible, but his proposed mechanism is not new, is considerably more radical than the phrase tax productivity suggests, and closely resembles a family of robot-tax proposals that tax scholars and economists have already spent years examining. The difficulty is not merely that it redistributes some of the gains from AI. It turns the advantage from becoming more productive into the object being taxed.
From wages to tokens
Gilan's July essay, Felkoden för en människa, is useful because he does not stop at the slogan. He attempts to describe the mechanism. For an established company, the proposal is reasonably straightforward: if a wage cost disappears and is replaced by a much smaller AI expense, a new levy captures some of the margin between them. Gilan describes this as the value created when lön blir tokens, when wages become tokens.
Suppose an activity previously costs SEK 1 million in labor and can subsequently be performed using SEK 50,000 of AI inference. The economically relevant margin under this proposal is approximately SEK 950,000, and some unspecified fraction of that margin would receive a special levy. This is therefore not simply a tax on purchasing tokens, nor is it equivalent to increasing VAT on an OpenAI invoice. Gilan explicitly rejects simply taxing inputs such as electricity because identical inputs can generate very different amounts of value. His target is the economic margin produced when expensive human labor is replaced by cheaper machine execution.
The difficult case is an AI-native company. No employee was dismissed, no historical payroll disappeared, and perhaps two humans with a collection of agents perform work that would traditionally have required a much larger organization. Gilan explicitly says in the same essay that such a company should still be subject to the levy. Since there is no historical payroll to measure, he proposes estimating what the work would have cost in human hours and translating that hypothetical labor into money. At that point the tax system is no longer observing an economic transaction; it is constructing a counterfactual production system and taxing against it.
That distinction is central to the entire problem. Actual wages, profits, consumption, dividends and capital gains are economic events that occurred. A hypothetical payroll describing how many humans might have been required under another production system is not an economic event. It is an administrative estimate of a world that did not exist.
The metric does not exist
Gilan points to METR's task-completion time horizons as evidence that human-equivalent work can already be measured. METR does indeed benchmark AI capabilities using human completion times, and the improvement has been striking, but METR itself explicitly warns against interpreting those measurements as ordinary professional work-hours. Its tasks are primarily self-contained software engineering, machine-learning and cybersecurity problems with relatively clear success criteria, while human baselines often come from skilled people entering with little project context. METR states that an eight-hour time horizon does not mean an AI can perform eight hours of a professional's normal job.
Real professional work contains tacit knowledge, accumulated context, organizational relationships, ambiguous objectives, incomplete evidence and outcomes that cannot always be scored by an automated evaluator. The distinction becomes even larger in productive AI use because the human operator may not be performing the same work as the machine at all. The scarce human contribution may instead be deciding what should be done, defining constraints, recognizing bad evidence, sequencing work, identifying when an answer is good enough and accepting responsibility for the resulting decision.
Human-equivalent execution time is therefore not equivalent to displaced payroll, economic value or taxable income. It is a capability benchmark being asked to perform the work of an accounting system, and converting it into kronor does not solve that category error.
The robot tax that is not called a robot tax
Gilan is careful to say that his proposal is not a "clumsy robot tax on physical machines." Literally, that is true. Economically, the distinction is much weaker because robot tax has never meant only a registration fee attached to an industrial robot arm. The term has long been used more broadly for tax instruments whose liability arises because automation substitutes for human labor. Once software or artificial intelligence performs the substitution, it belongs to the same economic family regardless of whether the machine has wheels, arms or an API.
The closest predecessor to Gilan's proposal is remarkably direct. In 2017, University of Geneva tax-law professor Xavier Oberson proposed taxing robots using an imputed hypothetical salary corresponding to equivalent work performed by humans. The tax would initially fall on the owner or employer using the machine and could also be associated with social-security contributions. The structure is almost the same as Gilan's AI-native mechanism: estimate the human labor that supposedly would otherwise have been required, construct a hypothetical salary from it, and derive a special tax liability from that amount.
This family of ideas was not ignored by scholarship. Christina Dimitropoulou's 2024 doctoral study, Robot Taxation: A Normative Tax Policy Analysis, gives extensive treatment to precisely these proposals. Her analysis includes a "conceptual attack" on applying imputed income to robots substituting for labor, a discussion of the inadequacy of deemed salary income, problems determining the deemed salary tax base, and what she calls false analogies between robot costs and prepaid labor services. Her conclusion is explicit: a robot tax based on hypothetical salaries of substituted labor is inappropriate, and the underlying policy problem is better treated as a reason to reconsider business-income taxation or the overall tax mix.
This does not establish that every conceivable automation tax is wrong. It does establish that the specific problem of inventing a human salary for a machine has already been identified, analyzed and found structurally difficult. The uncomfortable conclusion is therefore not that Gilan has proposed something unprecedented, but that he has independently arrived at a recognizable old instrument without engaging with the literature that already describes its failure modes.
The better operator problem
Consider a more contemporary production model. A senior technical adviser normally has an obvious scaling constraint: high-quality strategic work consumes attention, research time, analysis and preparation, and one person can only seriously serve a limited number of clients at once. The conventional responses are to raise the hourly price, hire more consultants, build a traditional firm, or simply accept the capacity limit.
AI introduces another possibility. Instead of charging five times as much, the adviser can maintain approximately the same market price while supporting materially more clients at once. Machine systems perform research, exploration, implementation, comparison and documentation in parallel, while the human remains responsible for framing the problem, determining what evidence is trustworthy, recognizing hidden risk, deciding what constitutes an acceptable answer and making the actual recommendation.
This is not only a theoretical future model. One senior technical operator I know already works this way and estimates that the practical leverage is roughly two to three times what was possible before current agentic AI systems became useful: not because the work is charged at two or three times the price, but because substantially more serious technical work can be kept moving concurrently. That is an anecdote rather than controlled evidence, and it should be treated as such. The interesting part is that the observed gain comes before anything resembling full automation of the professional role. A fivefold operating model is therefore better understood as a plausible direction to test than as a demonstrated productivity number.
Now compare two advisers with access to exactly the same models and approximately the same inference budget. One obtains a modest productivity improvement. The other has twenty-five years of technical experience, excellent domain models, strong judgment, good verification practices and an unusually effective way of decomposing and coordinating machine work, and therefore extracts several times as much useful productive capacity from the same underlying technology. The difference does not consist of synthetic senior consultants hiding somewhere inside a data center. A substantial part of it comes from the human operator.
As machine execution becomes cheaper, the scarce inputs can increasingly be judgment, experience, accumulated knowledge, orchestration ability and the willingness to accept responsibility for decisions. This is the production model described in The Centaur Manifest: cheap execution moves the bottleneck toward intent, verification, constraints and integration rather than eliminating the human contribution. A tax based on human-equivalent replacement work cannot separate the machine's contribution from the operator's ability to use it. The better the operator becomes, the larger the apparent quantity of replaced labor can become.
The special liability is therefore no longer merely attached to machine substitution; it is partly attached to human competence at exploiting the machine. Research on taxation and innovation gives good reason to be cautious about exactly this margin. Ufuk Akcigit, John Grigsby, Tom Nicholas and Stefanie Stantcheva found that personal and corporate taxation affects both the quantity and geographic location of innovation, while Akcigit and Stantcheva's broader review of taxation and innovation emphasizes innovation as a central source of long-run growth. These results do not imply that successful innovators should be tax-exempt; they show that expected returns influence investment in uncertain and potentially valuable activity.
The downside of experimentation is already private and real. Time can be wasted, capital can be lost, products can fail, reputations can be damaged and businesses can disappear. If exceptional upside receives an additional tax specifically because the experiment produced exceptional leverage, the risk-reward calculation changes. A system intended to encourage automation would then penalize the operators who become best at it.
A contradiction in the proposed instrument
Gilan correctly rejects paying companies to preserve jobs that machines can perform better and more cheaply. Such subsidies would reward inefficient production and attempt to freeze yesterday's production system in place. The productivity levy, however, introduces a closely related distortion from the opposite direction because the object being taxed grows with the economic advantage produced by automation.
If one company obtains a small advantage from machine execution, its automation margin is small. If another company discovers a radically superior production method, the difference between hypothetical human cost and actual machine cost becomes much larger. The policy therefore combines two incompatible propositions: firms should not be rewarded for remaining inefficient, but unusually successful efficiency gains should create an unusually large special tax base.
There is another problem underneath that contradiction. Productivity does not equal value. Producing five times as many reports does not create five times as much value, generating one million lines of code does not create more value than generating ten thousand, and an adviser running twenty agents can create enormous amounts of output while still making a poor decision. In the Outcome-Based Agile Framework, this distinction is explicit: an output is delivered, while an outcome is an observable change in reality. A human-hours tax risks treating theoretical execution capacity as a proxy for economic value even when the relationship is weak or nonexistent.
The spinning jenny problem
Historical automation provides a useful stress test because the same counterfactual problem appears long before artificial intelligence. In The Industrial Revolution in Miniature: The Spinning Jenny in Britain, France, and India, Robert Allen used the spinning jenny to explain part of Britain's early Industrial Revolution, arguing that unusually high British wages relative to capital costs made labor-saving machinery more attractive there than in countries such as France and India.
The revealing part for the present discussion is what happened when economists tried to calculate the machine's actual labor-saving advantage. Allen's original calculation effectively held output constant: if the jenny multiplied spinning productivity, a worker could produce the old quantity in a fraction of the previous time. In The Spinning Jenny and the Industrial Revolution: A Reappraisal, Ugo Gragnolati, Andrea Moschella and Emanuele Pugliese challenged that counterfactual. Why assume the worker would produce yesterday's quantity and then stop? A worker could instead continue working and produce substantially more yarn, changing the profitability calculation materially.
This is exactly the difficulty with hypothetical human-hours taxation. The old production function does not remain fixed while technology changes one number inside it. Lower costs change output volumes, prices, demand, organization, market size and which products become economical to produce in the first place. If an AI-native company can profitably sell a service for SEK 100,000 that would hypothetically have required SEK 2 million of conventional consultant labor, the SEK 2 million is not necessarily value that was displaced. The transaction might simply never have existed at the old cost.
A hypothetical payroll can therefore describe a world that would not have produced the product, company or market being taxed. The further technological production moves away from the old human process, the less meaningful the counterfactual becomes.
Industrial robots displaced workers and created productivity
Modern industrial robots provide better empirical evidence than eighteenth-century spinning, and the result is more complicated than either side of the political argument usually suggests. Daron Acemoglu and Pascual Restrepo's Robots and Jobs: Evidence from US Labor Markets finds real negative local labor-market effects in the United States: greater industrial robot exposure reduced employment and wages in affected commuting zones. That is strong evidence against the comforting claim that automation never causes meaningful displacement.
At the same time, displacement in one part of the economy is not a complete description of the economic effect. Georg Graetz and Guy Michaels' Robots at Work studied robot adoption across seventeen countries and found that robots contributed materially to labor-productivity growth, increased total factor productivity and lowered output prices. They did not find a significant reduction in aggregate employment, although lower-skilled workers lost employment share. Counting the workers removed from particular production processes therefore does not tell us the net economic effect.
European evidence on automation and taxation is similarly heterogeneous. Kerstin Hötte, Angelos Theodorakopoulos and Pantelis Koutroumpis' Automation and Taxation, covering nineteen EU countries from 1995 to 2016, found that the effect on tax revenues and tax composition changed across technologies and diffusion phases. The authors explicitly warn that concluding automation simply erodes the tax base can be myopic when broader technological responses are ignored.
The simple accounting identity -- one worker disappeared, therefore one worker's tax contribution disappeared -- captures only the first-order effect. Automation changes costs; costs affect profits and prices; prices affect demand; investment changes; complementary activities expand or contract; workers move; new firms become viable; capital income changes; consumption changes; and governments receive revenue from all of those flows. The net fiscal result is therefore an empirical question rather than something that can be inferred from the missing payslip alone.
Serious economics does not require preserving the old production function
There is a respectable economic case for automation taxes under particular circumstances. Daron Acemoglu, Andrea Manera and Pascual Restrepo argue that tax systems can themselves create excessive automation when labor is taxed substantially more heavily than equipment and software, allowing firms to choose machines partly because the tax system artificially changes the relative price of the two production factors. Their argument is fundamentally about tax neutrality rather than a moral preference for humans or machines.
That distinction is important. If automation is genuinely more productive, it should win. If human labor is genuinely superior, it should win. Neither should win simply because tax policy artificially changes the comparison. Correcting an existing distortion is conceptually very different from creating a new tax whose liability grows with the measured success of automation.
The IMF's 2024 Staff Discussion Note on generative AI and fiscal policy reaches a similarly cautious conclusion. Its baseline implication is that there should be no special tax on generative AI, robots or other labor-replacing technologies. It acknowledges that temporary automation taxes can sometimes be justified when labor-market adjustment is unusually slow and costly, but also notes that the optimal sign can vary and that under some circumstances the correct instrument can even be a subsidy rather than a tax. The same analysis warns that a special AI tax is difficult to define, can encourage production to move abroad and risks suppressing productivity growth.
Swedish economists Spencer Bastani and Daniel Waldenström arrive at a related position in their review of AI, automation and tax policy. They argue that a significant future increase in capital's share of national income could justify shifting the balance of taxation from labor toward capital, while simultaneously warning against damaging incentives for growth-enhancing entrepreneurship, innovation and investment in AI and automation. This is a considerably more cautious formulation of the same underlying fiscal concern.
The useful conclusion is not that the current Swedish tax system should remain unchanged forever. It is that if national income actually moves from labor toward capital, the tax system can follow the income that actually appears. No fictional workers are required.
Tax what happened
A useful tax base has a simple property: something economically observable occurred. Revenue happened, profit happened, a dividend happened, consumption happened, a capital gain was realized, land has a value, or a monopoly generated an economic rent. These are imperfect objects for taxation in different ways, but they refer to events and assets that exist in the actual economy.
By contrast, a statement such as "without AI, this work would have required 14.6 senior consultants for 1,820 hours" describes no transaction that occurred. It may be a useful business estimate, benchmark or thought experiment, but it is still a counterfactual. Different firms, different operators and different assumptions about output would generate different numbers, and technological progress gradually makes the comparison less meaningful rather than more precise.
An alternative tax system therefore does not have to choose between encouraging AI and financing the state. It can remain neutral about the production method and tax actual economic flows after they emerge. The OECD's Tax Policy Reform and Economic Growth documents how different tax instruments impose different costs on economic growth, with corporate income taxation generally more harmful to investment and long-run GDP per capita than broad consumption and recurrent immovable-property taxation. The exact Swedish tax mix is a separate political and technical problem, but the broader principle is useful: when taxation is necessary, the instrument should be chosen partly according to how little productive behavior it destroys.
If AI eventually generates very large economic rents in dominant firms, those rents are also conceptually different from productivity itself. A monopoly return protected by a bottleneck, network effect, exclusive resource or regulatory privilege is not equivalent to a skilled operator becoming unusually effective at applying models that competitors can also purchase. Treating both as "AI productivity" collapses fundamentally different economic phenomena into one tax base.
Norway owned an asset
Gilan proposes placing proceeds from the productivity levy into something resembling Norway's sovereign wealth fund. The destination may be reasonable, but the analogy does not justify the instrument that supplies the money. As the Norwegian Ministry of Finance explains, Norway converted revenues associated with an actual national resource endowment into financial assets. Those petroleum revenues arise from real taxation, resource ownership and direct state participation in petroleum activity.
Norway therefore had a scarce physical resource, production rights and actual economic rents from extracting that resource, and chose to convert part of those temporary flows into a diversified financial endowment. It did not estimate how many hypothetical workers the petroleum sector had replaced and levy a tax on their fictional salaries. If Sweden wanted a genuinely analogous AI strategy, the closer comparison would involve ownership of productive capital: public investment in infrastructure or diversified financial assets producing actual returns. Whether that is desirable is a separate political question, but the economic chain would at least remain coherent: productive asset, actual return, accumulated public wealth.
A tax on counterfactual labor followed by investment of the proceeds is simply a tax followed by a fund. The existence of the fund says nothing about whether the tax instrument feeding it is economically sound.
Human hours are becoming a worse model of work
There is a broader issue underneath the tax debate. For much of the industrial era, institutions have treated paid human labor as the canonical form of productive participation. Education prepares people for occupations, procurement frequently buys hours, employers count headcount and tax systems attach substantial revenue collection to payroll. The arrangement made sense in an economy where human labor was one of the dominant variable inputs in production.
Entrepreneurship has never fitted this model particularly well. An entrepreneur can spend years producing no return, make one unusually good decision and create enormous value. A senior engineer can solve in an hour a problem that consumes a less experienced team for weeks, while an investor can make a valuable decision in minutes because the relevant input was decades of accumulated judgment rather than the minutes consumed by the transaction. None of these activities has a stable relationship between hours spent and economic value produced.
AI amplifies that old discrepancy. The useful production equation increasingly resembles judgment + knowledge + ownership + risk + capital + machine leverage → economic outcome, rather than hours × effort → value. Hard work remains useful, sometimes indispensable, but it is an input rather than a unit of value.
Education will have to adapt if machine execution continues becoming cheaper. Preparing people merely to perform a defined occupation diligently is insufficient preparation for an economy in which comparative advantage can move quickly. People need to understand how to acquire new skills, evaluate changing markets, use leverage, own productive assets, tolerate uncertainty and recognize when reality has invalidated a previous strategy. There is no reason to expect every person to become an exceptional AI operator, just as there has never been reason to expect every person to become an exceptional entrepreneur, engineer, surgeon or musician. Human capabilities differ, and scarcity is precisely why some capabilities receive higher prices.
Flattening those differences is not a prerequisite for prosperity. A system that systematically reduces the return to unusually productive judgment can instead suppress the signal telling others which capabilities have become valuable.
An economy needs error signals
There is an important distinction between suffering and risk. Suffering is not productive in itself; willingness to accept risk is. The willingness to place time, capital, reputation or career opportunity behind an uncertain proposition is one of the mechanisms through which new production systems are discovered, and most attempts will not create exceptional returns. Some will fail completely.
The possibility of disproportionate upside is therefore part of what makes accepting the downside rational. Markets also require failure signals. If an activity no longer creates enough value to support its previous price, preserving the old reward suppresses useful information. If another production method creates several times the value from the same resources, suppressing its superior return removes useful information from the other direction. Continuous improvement requires visible differences between successful and unsuccessful strategies.
This is why policies intended to preserve particular jobs or historic income relationships become dangerous when technological production changes rapidly. The job is not the thing worth preserving; productive capacity is. Once another production method becomes genuinely superior, society becomes richer by allowing resources to move toward it, even though the resulting adjustment will not make every individual trajectory equal.
A productive system therefore needs room for people to take risks, fail, learn, outperform and receive disproportionate returns when they create disproportionate value. Removing every negative consequence would remove part of the feedback mechanism that produces adaptation; specially taxing exceptional productivity creates the same problem at the successful end of the distribution.
The diagnosis without the tax
Gilan has identified a serious possible future problem. Sweden cannot indefinitely finance a large state through human labor taxation if human labor eventually ceases to receive a comparable share of national income. A tax system designed around one production regime can become inappropriate under another, and there is no reason to assume that the relationship between employment, wages and national income will remain unchanged if machine intelligence continues improving.
That future has not yet arrived, however, and even if it does, the conclusion does not follow that productivity itself should become the tax base. Historical mechanization produced effects far more complicated than the disappearance of individual wage bills. The theoretical literature on automation taxation is conditional rather than categorical. The tax literature has already examined hypothetical robot salaries and identified the conceptual problems. METR's human-time metric does not measure what the proposed instrument requires. In the AI-native case, the mechanism ultimately requires the state to estimate how inefficient a company would have been in a world where its production technology did not exist.
A more defensible direction is less dramatic. Let productive technologies compete, remove tax distortions that artificially favor either humans or machines, allow unusually capable operators and firms to receive unusually high returns when they create unusually high value, and keep markets open enough that those returns attract competitors rather than becoming protected rents. If national income actually migrates from labor toward capital, taxation can gradually move toward the actual capital income and economic rents that emerge, using instruments selected for their economic effects rather than their political symbolism.
There is no need to reconstruct the workers who disappeared, and even less reason to reconstruct workers who never existed. If artificial intelligence eventually allows one capable human to coordinate the productive capacity that once required five, ten or fifty people, that is not a fiscal defect to be corrected. It is the productive breakthrough the technology was supposed to deliver, and the resulting economy should be designed to reward the people capable of turning cheap machine execution into valuable human outcomes rather than making them financially resemble the production system they replaced.
Not by shrinking man, but by enlarging what he guides.