Weighing the Evidence

The Stark Present: What Generative AI Is Actually Doing to Wages, Ladders, and the Tax Base

James Wolstencroft · July 2026

Lux Veritatis · Phronesis · Paper II of III

Abstract

Testis Temporum established what to look for: whether a labour-replacing technology is creating labour-enabling work fast enough to offset what it destroys, and whether the people being displaced have anything resembling the institutional bargaining power that eventually ended Engels’ Pause. This paper goes looking. It finds generative AI to be something narrower, and in its own way more dangerous, than the indiscriminate job destroyer either side of the public argument keeps describing: a skill compressor that lifts novices and flattens experts within the same occupation, while quietly starving the entry-level rung that produces the next generation of experts at all. It follows the same mechanism into software engineering, where it has converted the scarcest, most senior people in the building into full-time reviewers of their own juniors’ machine-generated work. And it follows the money into the specific fiscal architecture of the United Kingdom, a state that draws well over half its revenue from taxing exactly the kind of labour this technology is best at replacing, and asks what happens to that state when the taxable people start disappearing faster than the tax base can be redesigned. Whether anything can be done about it, and what doing something would actually require of a government, is left to the paper that follows this one.

1. The skill compressor, not the job destroyer

Generative AI compresses the range between a novice and an expert, inside the same job, and that is a much stranger thing to have built than either side of the public debate seems to have noticed.

Randomised controlled trials in customer support (Brynjolfsson, Li and Raymond) and in professional writing tasks (Noy and Zhang) land on the same headline figure: a 14% to 15% average productivity gain from generative AI assistance. Quoted alone, that number sounds modest, sensible, the kind of efficiency gain any tool should deliver. It is also misleading, in a specific way: averaging across a skill distribution hides the fact that the distribution changed shape, not just position. Break the same studies down by experience and the average dissolves into two different stories. Less experienced, lower skilled workers improve the speed and quality of their output by 30% to 34%. Highly skilled, experienced workers get almost nothing, and in some studies a small decline, because the model’s suggestions, tuned to a generic best practice, actively interrupt an expert’s already optimised judgement.

The mechanism is worth sitting with, because it explains why this is compression rather than simple assistance. Generative AI works by capturing the implicit best practices of high performing humans across an enormous dataset and handing them, on demand, to whoever asks. A junior worker with 2 months of tenure and a model behind them now performs at a level that used to take 6 months to reach unassisted. That sounds like a democratising effect, and in a narrow sense it is one: the floor rises. But a rising floor under a stationary ceiling is exactly what compression means. A wage structure built on the premise that experience is scarce, and therefore valuable, has just had its central assumption quietly removed from underneath it. Nobody voted on that.

2. The rung that stops being built

If a firm can buy junior-level output from a machine, it stops training junior humans to produce it, and the rung that used to turn juniors into seniors simply stops being built.

This is where compression turns into something closer to what Testis Temporum described in the handloom weavers’ workshops, not a factory floor being emptied all at once, but a specific, skilled pathway into a trade being quietly dismantled. Enterprise hiring data across firms adopting AI shows junior hiring, specifically the 22 to 25 age bracket, down by roughly 13%. That is the traditional entry-level rung of the career ladder being removed from underneath an entire cohort, not one bad year for graduate recruitment. It should worry anyone who has ever asked where next decade’s senior experts are meant to come from, because the honest answer, right now, is nowhere in particular. You cannot buy tenure. You can only let people accumulate it, on the job, making the small mistakes that teach them not to make the large ones, and that accumulation requires someone to employ them while they’re doing it. Sever the rung and you have quietly liquidated an asset the whole profession depends on 10 years from now, then booked the proceeds as this quarter’s profit and called it a saving on junior salaries.

3. The senior reviewer bottleneck

Software engineering is the clearest case study available, because it is the one industry that measures itself obsessively, and the numbers it has produced are damning.

An individual developer working with an AI coding assistant genuinely does feel faster, and task completion rates rise by roughly 26%. That is a real gain, worth saying plainly rather than burying under the caveat that follows. The caveat: individual speed and system speed are two different things, and a pipeline moves only as fast as its slowest stage, an old idea that AI has made newly, expensively visible. Because AI removes the historical constraint on how quickly code gets written, the constraint simply moves downstream, to review and integration. That is precisely what a longitudinal analysis of over 211 million lines of enterprise code found: duplicated code blocks up eightfold, code churn (code rewritten or deleted within weeks of being written) up ninefold, and active refactoring down by nearly 40%. Pull request volume is up 98%. Review time is up 91%. The code that arrives for review is, on average, 154% larger than it used to be.

Somebody has to catch what all of that generates, and the people catching it are, by construction, the scarcest and most expensive engineers in the building, because they are the only ones with enough architectural context to spot a subtly wrong assumption buried inside syntactically perfect code. Senior engineers, hired and trained to be strategic system architects, are being converted in practice into exhausted syntax validators, some reporting up to 66% of their working time spent debugging output that is almost right. Almost right is a uniquely draining category of error, worse in some ways than obviously wrong: the obviously wrong gets rejected in seconds, whilst the almost right has to be read in full before you find out which half of it is lying to you. The organisation ends up with more code, produced faster, reviewed more slowly, by fewer people qualified to review it. The debt this accumulates shows up on no balance sheet until the system it is buried inside fails in production, at which point it shows up as an emergency rather than a line item.

4. Britain’s specific exposure

Every advanced economy taxes labour more heavily than capital. Britain, more than most, has built its entire fiscal architecture on the assumption that this arrangement is permanent.

In the 2025/26 tax year, UK public sector receipts run to roughly £1.232 trillion, about 40% of GDP. Look at where that money actually comes from and the exposure becomes obvious. Income Tax raises £330 billion (26.8% of the total). National Insurance Contributions raise a further £204 billion (16.6%). Add Value Added Tax at £183 billion (14.9%), itself substantially a tax on the spending of people earning wages, and well over half the state’s income is drawn, directly or once removed, from human employment. Corporation Tax, the closest thing the system has to a tax on capital rather than labour, contributes somewhere between 8% and 10% (roughly £103 billion), a rounding error next to what labour is asked to carry. The system is also sharply progressive at the top: the top 10% of taxpayers contribute over 60% of total income tax revenue, the top 1% over 30%. That concentration is usually discussed as a fairness question. It is also a concentration risk, the part that gets left out of the fairness conversation: the revenue base depends disproportionately on a relatively small number of highly paid knowledge workers continuing to be highly paid knowledge employees, in exactly the category of work Section 1 has just shown compresses first.

The tax code does not merely fail to correct for this. It actively rewards it. Effective tax rates on human labour (payroll taxes and income tax combined) run at 25% to 30%. Effective tax rates on the software and hardware capital that replaces that labour run at under 5%, thanks to accelerated depreciation allowances and aggressive write-offs that predate generative AI by decades and were never designed with it in mind. Acemoglu, Manera and Restrepo have a name for what this asymmetry produces: so-so automation, technology adopted not because it is meaningfully more productive than the human it replaces, but because the tax code makes replacing the human profitable regardless. A firm facing a 25 point tax advantage for automating does not need the automation to be good. It only needs it to be cheap enough, after tax, to beat a salary, and that is a laughably low bar to clear.

5. When the taxpayer becomes the claimant

Every one of these mechanisms (skill compression, the severed ladder, the senior bottleneck, the capital tax advantage) points the same direction: fewer taxable wages, more benefit claims, at the exact moment the welfare state built to absorb the shock is itself funded by the wages disappearing.

This is where the proverb stops being a history lesson and starts being a live diagnosis. When a displaced worker stops earning, they do not simply exit the fiscal picture; they reappear on the other side of the ledger, drawing on Universal Credit and placing new demand on the NHS, whose own funding depends on the same labour tax base now shrinking. The Luddites and the Captain Swing rioters were reacting to exactly this dynamic in miniature: wages and dignity withdrawn at once, with no institutional channel to press the resulting grievance except the one they built with their own hands. The state met that grievance with troops rather than reform. Nobody is proposing hanging machine-breakers in 2026. But the underlying mechanics Testis Temporum named (an institutional arrangement that lets technology’s owners capture the surplus whilst its casualties absorb the cost) describe a tax code currently in force, not a period detail of the 1810s. And the description gets more accurate with every point so-so automation gains.

Call this what it is, plainly, because naming it is the only thing this paper is entitled to do before handing the problem to the one that follows. Britain in 2026 is rerunning the specific institutional failure that caused Engels’ Pause the first time: a labour-replacing technology deployed inside a tax and welfare architecture built for a different kind of economy, with no compensating shift yet made in who captures the gains. Nothing about that is coincidence. The first Pause took roughly 80 years to close, and it closed only once both the technology’s character and the institutions around it changed together. Whether either of those two things is currently changing, and what would need to be true for them to change on purpose rather than by accident, is the entire subject of Magistra Vitae.

Limitations and open problems

Three honest qualifications belong here, not buried in a footnote. First, “so-so automation” is Acemoglu, Manera and Restrepo’s diagnosis, not an uncontested consensus; other economists argue current AI deployment, code assistants especially, delivers genuine productivity gains large enough to justify the tax treatment on efficiency grounds alone, and this paper has not attempted to adjudicate that dispute, only to show that the tax asymmetry exists regardless of which side is right. Second, I use these tools daily, gladly, and they are genuinely good at what they do; nothing in this paper should be read as an argument against generative AI’s usefulness, only against the specific institutional arrangement currently deciding who captures the value it creates. Third, this paper stops at diagnosis by design. It does not evaluate a robot tax, a universal basic income, or any other remedy; that evaluation, robot tax included, belongs entirely to Magistra Vitae, and a reader who finishes this paper wanting a solution has one more paper to read before they’re entitled to one.

Boundary notes

Lux Veritatis takes the two questions Testis Temporum posed (is the technology labour-enabling fast enough, do the displaced have bargaining power) and answers both with present-day data rather than historical analogy: no, not yet, and no, not currently. It does not evaluate policy responses, which are the entire subject of Magistra Vitae, and it does not re-argue the historical precedent, which Testis Temporum has already established and this paper simply assumes.

References

• Brynjolfsson, E., Li, D., and Raymond, L. R. “Generative AI at Work.” NBER Working Paper 31161.
• Noy, S. and Zhang, W. “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence.”
• Acemoglu, D., Manera, A., and Restrepo, P. “Does the US Tax Code Favor Automation?” Brookings Papers on Economic Activity.
• GitClear. “Measure AI ROI with Research-Backed Developer Productivity Metrics.”
• Institute for Fiscal Studies. “Where does the government get its money?”; “How high are our taxes, and where does the money come from?”
• HMRC. “Tax receipts and National Insurance contributions for the UK,” annual bulletin.
• Office for Budget Responsibility. “Income tax and the earnings distribution”; “Tax by tax, spend by spend.”