Learning the Lesson

Being the Pupil, for Once: A Policy Architecture for the Age of Automated Labour

James Wolstencroft · July 2026

Magistra Vitae · Phronesis · Paper III of III

Abstract

Testis Temporum showed that Engels’ Pause closed only when two things changed together, the technology’s character and the institutions around it. Lux Veritatis showed that, in Britain today, neither has changed, and that the tax code is actively rewarding the version of automation least likely to change either one. This final paper is where the series stops diagnosing and starts prescribing, which is a much riskier thing to do in public and I’d rather admit that up front than pretend otherwise. It examines and rejects the most popular easy answer, a direct robot tax, on its own economic merits rather than on ideological grounds. It builds, instead, a settlement around four mechanisms: tax neutrality between capital and labour, a targeted levy on AI systems that directly replace a named human role, a sovereign wealth fund (fed by automation’s returns and by levies on the compute, energy and land it cannot offshore) that channels the proceeds back to the population that bore the cost, and an explicit, funded commitment to keep training the juniors the market has stopped paying to train. None of this closes the case on its own. Taken together, it is the closest thing this series has to an answer to the proverb it opened with: not a guarantee that the lesson gets learned, but a description of what learning it would actually require a government to do.

1. The easy answer, and why it doesn’t survive contact with software

A tax on robots sounds like justice arriving on schedule. It is also, on inspection, a policy that cannot be enforced against the thing it is supposed to tax.

The instinct behind a direct robot tax, popularised by figures including Bill Gates, is honourable, and it deserves to be stated plainly before it gets dismantled: if a machine takes a job, tax the machine at the rate the human would have paid, and use the proceeds to fund everyone else. It is simple, it sounds fair, and it fails for three specific, technical reasons that have nothing to do with anyone’s politics. First, software resists the very act of being counted. A robot on a factory floor is a discrete, countable object; an AI capability is a diffuse, scalable API call, embedded inside existing enterprise software, invoked billions of times, belonging to no single unit a tax inspector could point at and assess. Try to tax the unit and every vendor on earth immediately learns to bundle the capability into something that isn’t one, a tax arbitrage exercise that would take a competent accountant about a week to design. Second, a tax levied on the act of adopting technology is a tax on productivity itself, and a state that taxes the thing it needs more of (output, competitiveness, the means to fund the very services the tax is meant to protect) is cutting off the leg it’s trying to stand on. Third, and this one is almost philosophical, a great deal of what AI now does has no human predecessor at all; there is no historical wage to calculate an equivalent levy against, because no human was ever paid to do exactly that task in exactly that way. You cannot tax the robot at the rate the human would have earned when no human ever earned it.

So the robot tax fails on its own engineering, before its politics even get a look in: it isn’t really a coherent thing to build. Continuing to campaign for it wastes the one resource Lux Veritatis showed we are dangerously short of, reforming energy pointed at the right target. What follows aims at that target: the arrangement that rewards deploying automation for the wrong reasons, rather than the machine itself.

2. Ending the subsidy for so-so automation

If the tax code currently pays firms a 25 point bonus for replacing a human with a machine regardless of whether the machine is actually better, the first and cheapest fix is to stop paying the bonus.

Lux Veritatis established the numbers: effective tax rates on human labour running at 25% to 30%, effective tax rates on the software and hardware that replaces it under 5%. Closing that gap (raising effective rates on capital, easing them on lower and middle income labour) restores what economists call tax neutrality, and it does something more specific than sounding fair. It removes the artificial incentive for so-so automation, the deployment of technology that isn’t meaningfully more productive than the human it replaces but gets adopted anyway because the after-tax arithmetic favours it regardless. Firms would still automate wherever automation is genuinely, technically superior; nothing here proposes protecting inferior human labour from superior machines out of sentiment. What it removes is the free 25 points a firm currently collects for automating badly. A government that removes that free money is simply declining to keep paying for bad innovation out of its own revenue base, whatever the lobbying chooses to call it.

3. The Work Block: a harder lever, and its own honest tension

A more targeted proposal, the “Work Block” model, taxes an AI system at 100% of the salary of the specific human role it directly replaces, and it deserves to be presented with its own contradiction left visible rather than smoothed over.

Under this framework, when a firm deploys a system that replaces a named, previously human, job role, that role becomes a taxable AI Work Block, and the levy applied equals the full gross salary the human would have earned. A 100% tax sounds, at first read, like a policy designed to make automation impossible. It isn’t, and the reason is the same asymmetry Lux Veritatis diagnosed, run in reverse. The firm still saves the employer’s National Insurance contributions, pension matching, statutory sick pay, and the ordinary cost of turnover and absence; an AI system runs continuously, all year, with none of that overhead. So a 100% wage-equivalent tax still leaves automation profitable wherever it is genuinely efficient, whilst removing the subsidy that currently makes it profitable even when it isn’t. The revenue flows into the public fund Section 4 builds, so the substitution of a taxed human for an untaxed machine no longer defunds the state that both of them, in different ways, depend on.

I want to be honest about the tension this creates with Section 2, rather than letting two clean-sounding proposals sit side by side pretending they agree. They don’t, always, and smoothing that over is exactly the kind of tidiness that makes policy papers untrustworthy. Section 2 argues for equalising rates so the market decides, cleanly, where automation is genuinely superior. Section 3 argues for a targeted, heavier levy triggered by the specific, visible act of replacing a named human role. Push both too far simultaneously and you risk taxing genuinely beneficial automation twice (once through equalised capital rates, again through the Work Block), discouraging exactly the labour-enabling technology Testis Temporum showed ended Engels’ Pause the last time. The honest position is that Section 2 should be the default, general setting, and the Work Block should apply narrowly, to direct, named, one-for-one role replacement rather than to every use of AI in a workflow. That distinction will require better data on what “directly replaces a role” actually means than any tax authority currently collects. It is a real, unsolved measurement problem, not a rhetorical hedge, and I’d rather flag it here than let the mechanism sound tidier than it is.

4. A dividend, not a departmental budget line

Taxing automation better only helps if the proceeds reach the people automation displaced, and the cleanest mechanism for that, proposed by economists including Yanis Varoufakis, routes the money through ownership rather than through a means test.

A portion of corporate capital stock, IPO proceeds, and high-margin AI infrastructure profit is legally mandated into a public sovereign wealth fund, rather than into the ordinary discretionary budget, where it would compete every year against every other departmental priority and shrink the moment a chancellor needs headroom elsewhere. The returns on that fund, not the underlying capital, are distributed to every citizen as a Universal Basic Dividend. The mechanism has a specific virtue that a conventional tax-and-spend UBI doesn’t: it scales automatically. As capital’s share of national income rises (exactly the outcome Testis Temporum showed the Industrial Revolution produced for 80 straight years) the dividend rises with it, and no new act of Parliament is needed each time the automation share of the economy grows another point. It also sidesteps the reference-salary problem that sinks the robot tax outright, because a dividend from collective ownership never needs to know what a machine’s human predecessor used to earn. It only needs to know how profitable the machine turned out to be, which is a number firms already calculate for reasons of their own.

5. Taxing what a person cannot smuggle across a border

A UBI funded entirely by taxing human wages is a UBI that shrinks exactly when it’s needed most, so the base has to move to things that cannot leave the country the way a payroll line item can.

Compute usage, data centre capacity, and the specialised hardware AI runs on are physically anchored to a jurisdiction, at least for now, in a way a remote knowledge worker never is. Levy that usage, price the carbon in the energy it consumes, capture the value of the land under its footprint, and the state has a tax base that grows precisely as automation grows, rather than one that erodes with it. Each of these already exists in some form somewhere in the world. What’s missing is the decision to route them specifically toward the safety net that automation itself is straining, rather than treating them as general revenue indistinguishable from everything else the state collects. So, to be plain about the plumbing: three intakes, one fund, one dividend. The Work Block levy of Section 3, the capital returns of Section 4, and the anchored levies of this section all arrive at the same place, on purpose, because a single visible dividend is far harder to quietly raid than three scattered budget lines.

6. What none of this money fixes on its own

A cheque, however generous, does not replace what the Luddites were actually protesting, and a policy architecture that forgets this will fund the right amount of the wrong solution.

Testis Temporum was explicit that the grievance behind the Luddite uprisings was never simply about wages; it was about occupational autonomy, about a skilled trade being unbundled and handed to workers with none of the craftsman’s standing or control over how the work got done. Work provides social connection, structure, and a sense of one’s own competence, none of which a dividend, however well funded, delivers on its own. A UBI or UBD paired with nothing else risks repeating, in a softer register, the exact mistake nineteenth century governments made: treating a crisis of dignity and structure as though it were only a crisis of income, and buying quiet with a payment instead of addressing what actually broke. So the fourth mechanism is institutional rather than fiscal. Governments and the firms benefiting from automation’s gains have to fund, explicitly and permanently, the junior pipeline that Lux Veritatis showed the market has stopped paying for on its own: restructured career ladders that assume AI assistance rather than pretending it away, funded modern apprenticeships, and protected mentorship time for the senior engineers currently being converted into full-time reviewers. That protected time gets spent teaching the next generation rather than only catching the last generation’s mistakes. Income security buys the floor. It does not, on its own, buy back the ladder, and a government that funds only the floor has learned half the lesson and mistaken it for the whole thing.

7. The pupil, finally

Three papers ago I opened this series with an Austrian proverb about a classroom that stays empty no matter how many times the lesson gets taught. It would be a tidy piece of writing to end here by declaring the lesson finally learned. It hasn’t been, not yet, not in any parliament I’m aware of, and I’d be lying to you, and to myself, if I dressed this closing section up as a victory lap rather than what it actually is: an architecture for what learning would look like if a government ever decided to try. Testis Temporum showed the classroom has been open before, for 80 years, whilst the people inside it paid the tuition in wages that never moved. Lux Veritatis showed the same lesson is being taught again, right now, in numbers rather than in looms. This paper is the only thing a proverb about refusing to learn actually leaves room for: not a promise that the lesson lands this time, but a syllabus, fully written out, for the one government brave enough, or simply frightened enough, to finally take the seat.

Limitations and open problems

This paper is architecture, not legislation, and three gaps deserve naming rather than hiding. First, none of the four mechanisms proposed here has been modelled at full macroeconomic scale for the United Kingdom specifically; the fiscal feasibility literature this paper draws on, the Bath University costings among them, models variants of UBI, not this exact combination, and the interaction effects between a Work Block levy, equalised capital taxation, and a sovereign dividend paid from the same tax base have not, to my knowledge, been jointly modelled by anyone yet. Second, this paper has said almost nothing about political feasibility, which is not the same question as economic soundness, and a mechanism can be entirely correct on the page and entirely unpassable in a legislature that answers to voters on a five year cycle while capital’s share of income moves on its own, much longer, schedule. Third, the tension flagged in Section 3, between equalised capital taxation and a targeted Work Block levy, is real and currently unresolved; I have proposed a default and an exception rather than a proof that the two mechanisms compose cleanly, and Reasoning in the Fog, elsewhere in my own work, has more to say about exactly this kind of composition risk between individually sound mechanisms than this paper has room to repeat.

Boundary notes

Magistra Vitae assumes the precedent of Testis Temporum and the diagnosis of Lux Veritatis without re-arguing either. It is the only paper in this series that prescribes rather than describes, and its limitations section should be read as a genuine account of what remains undone, not as a formality appended out of academic habit. The question this paper leaves standing, what a government that has agreed to learn actually looks like, is the subject of Psephos, entered through its prologue, Setting the Keystone (forthcoming; unpublished at the time of writing).

References

• Gates, B., quoted in “Should we tax robots?” Policy Forum.
• Varoufakis, Y. “Taxing robots won’t work.” World Economic Forum.
• Acemoglu, D., Manera, A., and Restrepo, P. “Optimal policy in a task framework.” NBER Working Paper 27052.
• LSE Business Review. “Universal basic income as a new social contract for the age of AI.”
• University of Bath. “The Fiscal and Distributional Implications of Alternative Universal Basic Income Schemes in the UK.”
• McGill Law Journal. “I Robot: U Tax? Considering the Tax Policy Implications of Automation.”
• Brookings Institution. “Does the US Tax Code Favor Automation?”
• IBFD. “Taxing Artificial Intelligence and Robots: Critical Assessment of Potential Policy Solutions.”
• Lewis Silkin. “Robot Tax: the pros and cons of taxing robotic technology in the workplace.”
• Wolstencroft, J. Reasoning in the Fog (The Tekmerium Series).