I’ve been thinking about the tariff debate not as a trade policy question but as a symptom. The surface argument is about goods crossing borders. The actual argument underneath — the one nobody wants to state plainly — is about who funds the society that businesses sell into, and what happens to that question as automation removes the last obvious answer.

The stated purpose of a tariff is simple: make imported goods more expensive so consumers buy domestic ones instead. The political argument is that foreign countries are taking advantage of us, that our companies moved manufacturing abroad, that tariffs level the field.

That argument has real problems. Tariffs are blunt, they raise prices for domestic consumers, and they invite retaliation. Most economists will point you to better tools. There’s a second use — tariffs as geopolitical leverage, a tool to extract concessions or rewire supply chains toward allies rather than adversaries. That dimension is real, and it’s a legitimate policy instrument. It’s just not what this post is about.

But underneath the tariff intuition is a different complaint that doesn’t get stated cleanly, and that one is harder to dismiss. I’m not an economist, and this isn’t policy analysis — it’s an attempt to follow the thread honestly.

The Problem Tariffs Are Trying to Name

Here’s the actual grievance: a company can design its products in one country, manufacture them in another, sell them to consumers in a third, and route its profits through a fourth. The consumers in that third country provide all the revenue. Their society provides the roads, the legal system, the educated workforce, the currency stability that makes commerce possible. And yet the company’s obligation to that society — in taxes, in wages, in any form of contribution — can be structured to approach zero.

This isn’t hypothetical. Apple designs in California, manufactures in China, and historically routed profits through Ireland. That structure eventually triggered a €13 billion EU back-tax ruling — the EU found Ireland had given Apple illegally favorable tax treatment. The ruling was finally upheld by the ECJ in September 2024, after nearly a decade of litigation, with the money having sat in an Irish escrow account since 2018 pending appeal. That’s a real governance response. It also required a supranational court, took roughly a decade, applied only to past behavior, and changed nothing structurally about the framework that made it possible. Amazon automates warehouse fulfillment with robotics while still selling to American consumers and depending on American roads, courts, and payment infrastructure to operate. Both are operating legally, within frameworks that weren’t designed to handle this configuration.

Tariffs are an attempt to capture value at the point of transaction. Crude, but the underlying concern is legitimate: companies should contribute something to the societies they profit from.

The mechanism by which companies have historically made that contribution wasn’t designed — it evolved. When a company hired workers, those workers paid income taxes, contributed to Social Security, spent locally, and collectively funded the public infrastructure the company’s own operations depended on. Employment became the primary transmission mechanism between private enterprise and public goods. Not a tax, not a regulation — just what happened when a company needed humans to function.

The problem is that tariffs only address one version of this — goods crossing a border. They don’t touch the version that’s accelerating.

The Version That Doesn’t Cross a Border

A company doesn’t have to manufacture abroad to escape contributing to the country it sells into. It can automate.

A robotized warehouse doesn’t cross a border. Software replacing a call center doesn’t cross a border. An AI agent handling work that employed fifty people doesn’t cross a border. The company can be entirely domestic and still reduce its obligation to the society it profits from — because the primary mechanism by which companies have historically contributed (employing people who pay income taxes, spend locally, and fund public services) is exactly the mechanism being removed.

This is where the tariff conversation becomes inadequate. The policy tool is aimed at goods crossing borders. The structural shift is happening inside borders, in every industry, simultaneously.

The incentive alignment makes this structural rather than incidental. Corporate governance — reinforced by public market pressure and compensation structures tied to shareholder returns — creates systematic obligation to maximize returns to shareholders. Automation is one of the most reliable ways to do that: it converts recurring labor costs — which generate payroll and income tax obligations — into one-time capital expenditure, where accelerated depreciation provisions under the 2017 Tax Cuts and Jobs Act substantially reduce the effective tax cost (though bonus depreciation had been phasing down since 2023 before being restored to 100% for property placed in service after January 19, 2025 by legislation enacted that year). The legal structure, the tax code, and the technological capability all point the same direction. No current policy points the other way.

As automation extends to AI — and it is extending at an accelerating rate — the need for human workers shrinks across the economy. We’re watching it happen in knowledge work right now. Robotics will follow in physical industries. The pressure on employment, if current incentive structures hold, would exceed what existing safety nets are designed to absorb.

The historical counterargument is serious and deserves a direct response. Agricultural employment collapsed from 41% of the U.S. workforce in 1900 to 2% by 2000 (USDA Economic Research Service ) without causing mass unemployment — workers moved into manufacturing, then services, then knowledge work. Every prior automation wave triggered displacement fears that didn’t materialize at scale. Research formalizes why : automation removes tasks from humans, but technology consistently creates new tasks where humans retain comparative advantage. Other economists argue AI could even augment lower-skill workers by letting them perform tasks previously requiring expert credentials — AI as tool rather than substitute.

The argument for why this wave may be structurally different rests on three things. First, cognitive breadth: prior automation displaced physical labor or narrow repetitive tasks; generative AI spans legal analysis, software development, medical diagnosis, and creative work simultaneously — unprecedented scope. Second, the reinstatement effect itself appears to be weakening. A 2022 Econometrica study attributes 45% of industry labor share declines from 1987–2016 to automation, with new task creation failing to keep pace for lower-skill workers — the cushioning mechanism that made the historical prediction wrong may be underperforming. Third, there’s a category economists have called “so-so technologies” (a term from Acemoglu & Restrepo’s earlier automation literature that Acemoglu (2024) applies specifically to AI): automation that reduces costs without meaningfully raising productivity, where gains flow to capital rather than workers, breaking the wage-raising mechanism that historically funded new job sectors. This isn’t a projection — it’s a 50-year trend. From 1948 to 1973, productivity and median compensation grew in parallel. Since 1973, productivity has grown far faster than median wages — the Economic Policy Institute documents productivity growing 72% versus median hourly compensation growing 9% from 1973 to 2014 — with the gap flowing to capital owners rather than workers. The mechanism Acemoglu describes is the explanation: when productivity gains come from augmenting workers, competitive labor markets translate them to wages; when they come from replacing workers with capital, gains go to whoever owns the capital. Acemoglu (2024) models the macroeconomic implications of this dynamic, projecting that AI’s overall GDP impact will be substantially more limited than tech-optimist forecasts assume.

Current data supports a split outcome, not a binary. The Yale Budget Lab’s ongoing tracker finds no aggregate unemployment signal tied to AI exposure yet. But the Dallas Federal Reserve’s January 2026 paper found a 13% employment decline since 2022 among workers aged 20–24 in the most AI-exposed occupations — retail supervisors, administrative assistants, customer service representatives — driven by reduced new-hire inflows rather than layoffs. The people competing with AI face displacement pressure. Whether the reinstatement effect eventually catches up — as it has in prior waves — or whether the cognitive breadth of this wave makes the historical analogy break down, is the genuine empirical question. This post doesn’t assume we know the answer. The historical analogy may hold; the cognitive breadth argument may prove decisive. The governance case below doesn’t depend on which scenario plays out — it depends on recognizing that the window to design the structures is the same window as the transition, and that window is already open.

The Income Floor Problem

The pattern is already visible in the data. In December 2024, the CBO published its first-ever analysis of AI’s fiscal impact (CBO, Dec 2024 ), identifying automation-driven job displacement as a key fiscal risk — noting that permanently displaced workers would reduce income and payroll tax receipts while increasing safety-net claims. The report did not predict which scenario would prevail, presenting productivity gains and displacement as both plausible possibilities.

The standard counterargument here is deflationary: if automation drives down the cost of goods and services, workers may be better off in real terms even if nominal wages stagnate or fall. This is not a weak argument — manufacturing automation contributed to decades of goods price deflation, and AI is already collapsing the cost of software, customer service, and information work. But the deflationary benefit is uneven. It concentrates in tradeable goods and digital services — things that are already cheap or becoming free. The non-deflationary stickiness is in housing, healthcare, and education — exactly the categories where household costs have been compounding upward for decades and where automation has made the least inroad. Cheaper electronics don’t offset rising rent. The deflation and the wage pressure don’t cancel out.

In a displacement scenario, some form of income floor — whether called UBI or something else — starts looking less like a policy preference and more like a structural requirement for the consumer economy to function. If a large portion of the population has no income, the economy seizes. People have to be able to buy things for the companies producing those things to survive.

But UBI comes from government. And government funding is currently structured primarily around taxing individuals — income tax, payroll tax, consumption tax on wages spent. As those wages disappear, the tax base collapses. The government finds itself trying to fund an expanding safety net with a shrinking revenue source.

The gap has to come from somewhere. The logical somewhere is corporate taxation — closing the structures that allow companies to accumulate productivity gains without proportional contribution to the societies they operate in.

This is easy to say and hard to implement. The complexity is real: what does a company “owe” to a country when its supply chains span twenty countries and its customers are global? How do you calculate the obligation?

What a More Honest Framework Might Look Like

The current policy debate mostly reaches for a robot tax: charge companies for each position replaced by automation. The impulse is right — automation does remove the payroll-based contribution that historically funded public services. But the mechanism fails on first contact with implementation. Defining what counts as a “robot” is functionally impossible — does Excel qualify? A CRM that eliminated five data-entry roles? A language model handling work that employed a team? And taxing capital directly penalizes the productivity gains that make the disruption economically defensible in the first place. South Korea’s 2017 experience confirms this: a partial rollback of automation investment credits led to a 28 percent relative decline in robot installations among affected industries, with no statistically significant employment benefit. A third obstacle is jurisdictional: any unilateral robot tax creates immediate competitive disadvantage — the EU Parliament explicitly cited this when it rejected the proposal in 2017, noting that regions implementing it would lose ground as companies relocate operations. Other proposals — wealth taxes, sovereign wealth fund equity stakes in AI companies — address wealth concentration but don’t directly solve the contribution gap as automation scales. The standard economist answer — a consumption tax (VAT) rather than a production or employment tax — sidesteps the measurement problem entirely and doesn’t penalize automation; the tradeoff is that it shifts the burden toward consumers rather than corporations, which is a different political choice.

Acemoglu’s own prescription follows the same logic: rather than adding a new robot levy, address the existing subsidy — the bonus depreciation provisions that already tilt the tax code toward capital over labor. The framework below extends that reasoning: instead of measuring what gets automated, measure what gets contributed.

One way to think about it: tie the obligation to something concrete and measurable — the same logic that makes the individual tax system work.

Right now, the IRS doesn’t track every worker’s income directly. It distributes the obligation to employers, who report via W-2s and 1099s with Social Security numbers as the identifier. The system works because the identifiers are unambiguous and the enforcement is distributed — the IRS doesn’t need to watch you; it watches your employer, and your employer has every incentive to get it right.

A framework for the automation problem might use the same self-reporting model for businesses. The identifiers that would actually measure obligation:

  • Domestic payroll as a percentage of revenue — a W-2 ratio (adjusted for sector norms — capital-intensive industries naturally carry lower ratios). Companies that employ proportionally few people relative to their domestic revenue are consuming more public infrastructure per dollar of obligation than their payroll suggests.
  • Infrastructure consumption metrics — utilities, logistics networks, court system usage, regulatory protection consumed. Some of these are measurable; others — regulatory protection, legal system access — are harder to quantify but not impossible to proxy.
  • The gap between those two numbers is what current tax law doesn’t price.

The IRS parallel extends further: just as the income tax puts reporting on employers because they’re the organized party with auditable records, an obligation framework like this would put reporting on companies — not because they’d comply perfectly, but because the obligation would be legible and verifiable in a way that individual worker tracking never was.

This approach would sidestep the robot tax’s definitional problem entirely: it doesn’t try to measure automation. It measures the gap between what a company earns in a domestic market and what it contributes via payroll. That data already exists and is already reported.

One real vulnerability: a payroll/revenue ratio would create an incentive to minimize visible domestic headcount — through contractor reclassification, domestic outsourcing to staffing firms, or subsidiary restructuring. Any framework of this kind would need to define “obligation” broadly enough to capture 1099 relationships and layered employment structures, not just W-2 headcount. The IRS system works partly because employers have strong liability incentives to report accurately; a framework that inverts that incentive needs explicit design against it.

Would this be onerous to administer? Yes. Would it be more accurate than a border tax that misses the domestic automation version of the same problem entirely? Probably also yes.

An interim version of something like this might index tariff equivalents to domestic payroll percentage — companies with high domestic employment pay less, companies that have automated their domestic operations while retaining their domestic customer base pay more. Crude, but at least aimed at the right target.

One significant caveat: this framework is US-specific by design. The IRS mechanism, W-2 reporting, and payroll tax infrastructure are American law. The problem is global — a company can automate in Germany, sell into the US market, and the same contribution gap exists across multiple jurisdictions simultaneously. A domestic framework addresses part of the problem; closing it fully requires international coordination of the kind that has historically moved slowly and incompletely. The OECD’s global minimum corporate tax — agreed in principle in 2021, implemented patchily since — is the closest precedent for what that coordination looks like in practice.

Why Even This Breaks Down Eventually

Here’s the harder part: even a well-designed framework tied to employment doesn’t work indefinitely, because employment is the thing going to zero.

If the endpoint is that robots produce most goods and services, the cost of producing things approaches near-zero for anything that isn’t resource-constrained. Services — software, information, logistics, support — become effectively free to produce.

What remains expensive is physical resources: materials, energy, feedstock. That constraint may not be permanent, but the honest version of the argument is narrower than “costs approach zero.” The threshold that matters is break-even: once the cost of extracting resources — whether from remote regions or eventually from space — falls below the value of what’s returned, the loop becomes self-sustaining. You don’t need near-zero costs. You need retrieval to be worth doing, and then it funds more retrieval. Whether that crosses in decades or generations is genuinely uncertain. But it shifts physical resource scarcity from a hard ceiling to an engineering and economics problem — a different category of obstacle. Near-zero cost services. Resources whose effective cost keeps falling as the extraction loop scales. The theoretical endpoint is a society where the material constraints on human needs are substantially removed, though how far and how fast is an open question.

What Abundance Doesn’t Fix

The standard pushback on post-scarcity arguments is that they’re idealistic about human nature. The specific challenge is about abundance — the idea that post-scarcity would be good, or that the problem reduces to designing the right distribution mechanisms.

The challenge is fair. And I think the honest response is to grant most of it.

The assumption buried in the post-scarcity framing is that humans primarily want things — goods, comfort, security — and that competition and hierarchy are downstream of scarcity. Give everyone enough, the thinking goes, and you drain most of the pressure out of the system.

One reading of human nature — one I find persuasive, though it’s contested — is that humans don’t primarily want goods; they want status, and goods are one arena in which that competition plays out. On this view, the drive toward control and access to resources isn’t a consequence of scarcity — it’s more fundamental. Scarcity channels it; scarcity doesn’t cause it.

Which means removing material scarcity doesn’t defuse the drive. It just changes the arena.

A world of material abundance could easily be a world of intense social hierarchy, because hierarchy is what the competition was always actually about. The people who manage to control distribution in a post-scarcity world won’t be doing it because they need more things. They’ll be doing it because control itself is the point — and it always was.

The mechanisms are familiar. Vote a certain way, or your allocation is delayed. Support the right people, or find yourself at the back of the queue. The scarcity wasn’t abolished — it migrated, as it always does, to whatever ceiling remains. The people who control the distribution systems in a post-scarcity world have more power than the wealthiest individuals in a scarcity world, because the levers are social and political rather than economic, and those are much harder to route around.

So the governance problem isn’t just “how do we make sure the abundance gets distributed fairly?” That’s the tractable version of the question. The harder version: how do you design structures that channel the human drive toward control in directions that are less destructive? History has partial answers to that — not permanent solutions, just contested arrangements that have to be maintained against constant pressure from people trying to capture them. That’s a much older problem than anything in this post. But it’s why the design choices made during the transition matter — the mechanisms that constrain the drive toward control have to be built into whatever structures get established now, before the defaults harden.

The Window That Closes

The thread running through all of this: tariffs are a real-time response to a real-time trade dispute, but the underlying question — how do societies fund themselves when human labor stops being the primary input to production? — is what the tariff intuition is actually reaching for.

We’ve seen this pattern before. Nobody acted seriously on governance of social media platforms until after they had accumulated the scale and leverage to shape elections. The frameworks — antitrust doctrine, content moderation law, data privacy regulation — were improvised after the power was already concentrated. We’re still arguing about whether any of it works. The analogy isn’t perfect — AI’s labor impact is more actively contested in advance than social media harms were — but the structural pattern holds: the window to design frameworks is narrower than it looks.

The same risk applies here, at larger scale. The transition period — when automation is accelerating but hasn’t yet displaced the majority of labor — is the window when it’s still possible to design the funding and distribution structures in advance. The structures can be argued about, iterated on, adjusted based on what’s actually happening. That window requires political will that’s currently absorbed by a debate about border taxes.

Once the disruption completes, the question becomes: who already controls the robots? The governance structures that determine whether post-scarcity abundance gets broadly distributed, or whether it concentrates into a new form of control, don’t design themselves. They either get built during the transition — when there’s still room to negotiate them — or they get imposed after the fact by whoever got there first.

Power doesn’t voluntarily constrain itself. Antitrust law, Social Security, utility regulation — at best, they were improvised during the disruption rather than designed before it. They were built after the disruption was mature, the damage was legible, and the political cost of inaction finally exceeded the cost of doing something. That’s not a reassuring model. It’s a description of how expensive delay gets, and how partial the fixes are when they finally arrive.

Without deliberate design, the default outcome here is concentration — not as conspiracy, just as the natural result of first-mover advantage compounding without countervailing structure. The window to design something different is the same window as the transition.

The transition is already underway. The window is already open. The question is whether the political will to use it arrives before the window closes — or after, when the answer to “who controls the robots?” is already settled.

References

A note on how this was made: The core argument and thesis are mine — I came up with the central idea and framing. The post was written collaboratively with Claude (Anthropic’s AI), which helped shape the prose, identify and synthesize the research, provided citations that I verified against the source papers, and raised several questions I hadn’t considered — including the counterargument about the historical pattern of automation creating new jobs. That division of labor is the kind of workflow this blog explores.

Related: What are the actual odds — utopia, dystopia, or the muddle? | What can we actually do about it? | The cost of software is falling — what happens next?