Robots earn no wages. The people they replace still need to live.
A company can remove labor costs from its automation report. It cannot simultaneously remove household rent, community demand, or society's retraining and support costs.
In an earlier Zhihu answer, I called the tendency to count corporate savings while ignoring social adjustment delayed social liabilities. Liability here is a metaphor for costs excluded from decisions, not a claim that every automation creates an accounting liability.
Physical AI makes the question more concrete: as machines perform more real-world work, how do corporate efficiency gains connect with society's ability to pay?
A saving in one ledger may become an unpaid cost in another. That transfer is where a robot-tax discussion should begin.
But one fewer employee does not necessarily mean one permanently lost tax contribution. Higher profits, lower prices, expanding demand, and new jobs can change the outcome. Track where income and costs go before deciding where taxation and support should follow.
This essay continues my discussion in this Zhihu answer:How should we assess the claim that AI is creating a deeply unbalanced economic structure?（Read the full answer）。
Whose costs are missing from corporate ROI?
Imagine a company spending US$100 million on automation to save US$30 million a year in labor costs. This is an illustration, not a company case. Even its own return calculation must include maintenance, energy, financing, downtime, deployment, and workforce transition, rather than saved wages alone.
The social account adds questions. Wages a firm no longer pays should not be counted in full again as lost social resources: they may correspond to higher profits, lower prices, or other income changes. Unemployment benefits are fiscal spending and income transfers, and should not simply be added to real resource costs. Social appraisal must identify additional resource use, harms not reflected in private decisions, and distributional effects. Who pays for displaced workers' job search, training, and relocation? Does the local social-insurance contribution base shrink? How do households maintain spending after losing income? Excluding these costs from a project budget does not remove them.
My earlier warning that companies may weaken their own customer base concerns purchasing power. One firm's job cuts need not reduce its sales. But if many firms depress labor income and gains do not sufficiently circulate through lower prices, new jobs, investment, or redistribution, demand may suffer. Individually rational choices do not automatically produce a stable aggregate outcome.
This asks for fuller investment analysis, not opposition to automation. Removing dangerous work, augmenting workers, reducing prices, and broadening services have real value. Benefits and transition costs belong in the same picture.
Why is the robot tax back in the debate?
In an August 2026 essay, Bill Gates again advocated taxes on robots and AI tokens to fund transitions and social protection and moderate labor substitution. This policy proposal makes a clear value judgment: some loss of short-term efficiency may be worthwhile to reduce transition costs.Gates's essay
Saying technological progress should not be punished does not settle the question. Adoption's gains and costs may fall on different groups: a firm saves labor costs while workers and communities pay for job search, relocation, and skill depreciation. Public policy has reason to examine that account.
The problems with taxing machine counts
The difficult question is what counts as a robot.
Two production lines might achieve equal automation, one with humanoids and another with conveyors, fixed arms, and software. Taxing only the first encourages equipment redesign without necessarily changing labor substitution. Extending the definition to every labor-saving technology could capture self-checkout, scheduling software, and ordinary upgrades.
Machine counts also ignore capability. A robot working one hour daily differs economically from one running continuously. Token-based charges may not track value either: model usage is not profit or a measure of jobs replaced.
An apparently simple tax can therefore be hard to administer.
This IMF analysis advises against a specific AI tax while recommending review of incentives favoring labor substitution and adaptation of capital-income and corporate taxation. These positions are compatible: avoiding a tax on machine appearance need not preserve every existing subsidy.IMF policy analysis
Four accounts to track when income changes
I would separate the debate into four accounts.
First, labor income. What happens to total wages, hours, and the social-insurance contribution base? Is technology, the business cycle, or industrial relocation responsible? Lower taxes on low-paid work can be considered, with funding specified. A tax-cut proposal is incomplete unless it explains how public services will still be funded.
Second, capital income. Where are new profits recognized, who owns them, and can applicable taxes be collected? Distinguish normal returns compensating investment risk from excess returns arising from scarce resources or market barriers. Financial statements do not always separate them easily; policy must acknowledge identification and enforcement costs.
Third, consumption. Automation may increase real consumption, but lower prices alone do not ensure a larger consumption-related tax base. For sales taxes or VAT generally levied on transaction value, the result also depends on taxable nominal spending, tax rates, and exemptions. Higher consumption taxes can also burden low-income households. Assess taxes together with rebates, cash support, and public services to establish net household burdens, beyond nominal rates.
Fourth, transition spending. How much training, unemployment support, and local adjustment is needed, and when? A locality losing wage income this year does not automatically receive help because another region earns higher profits years later. Fiscal challenges include timing and geography as well as totals.
Korinek and Lockwood's Brookings framework emphasizes consumption taxes and economic rents in the AI era while distinguishing final consumption from business inputs. It is an approach worth comparing, not a ready-made answer for every country. Taxes on consumption, profits, and capital each raise distributional and implementation problems.Brookings public-finance framework
Can affected people afford to wait?
The strongest objection is that people cannot wait until researchers finish tracing income and tax incidence.
That concern is justified. More research cannot excuse indefinite delay. Early action can address clear needs: extend support to workers losing hours, fund verifiable training and job transitions, and offer time-limited, evaluated help to affected regions. These measures can proceed alongside tax reform without a perfect robot definition.
Account for policy risk without treating forecasts as law
My earlier answer argued that firms cannot assume society will absorb adjustment costs indefinitely. That remains true. But distributional pressure prompting policy change does not mean an AI tax is inevitable; institutional choices, legislation, and enforcement still intervene.
Governments could alter labor and capital taxation, broaden protection, support job transitions, or change investment incentives. A dedicated robot tax is one option. Outcomes depend on whether shareholders, consumers, or workers bear the burden and how revenues are used.
Businesses should not build their investment models around the assumption that today's rules will never change. Scenarios for different tax, employment, and support arrangements offer more than predicting every saved wage will return as a penalty.
Likewise, collaboration projects or services for disadvantaged groups do not automatically earn exemptions or policy protection. Their value should be demonstrated through improvements in safety, income, access, and work quality; actual treatment depends on applicable rules.
Start with fiscal stress tests, then discuss new taxes
Governments could publish automation fiscal stress tests: under different wage, profit, and employment scenarios, which revenues change, which expenditures rise, and how do local and central authorities share costs? Include productivity benefits, rather than modeling only lost jobs.
Businesses could add a transition budget to automation reviews: who needs training, who may lose hours, who pays, and do suitable jobs actually exist afterward? Funding verifiable transitions, job redesign, and development is more convincing than promising future jobs. Maintenance, operations, and exception handling may offer opportunities, but their numbers, locations, and skills need not match displaced work.
Public debate should require every tax proposal to answer who ultimately pays, how double taxation is avoided, and what results would trigger adjustment or withdrawal. A new name cannot replace those basics.
Distinguish short-term income stabilization from long-term tax-base reform. Temporary help can have deadlines and review conditions; permanent promises need stable funding. A robotics company's rising valuation cannot substitute for assessing society's capacity to pay, nor can possible future prosperity erase today's transition costs.
Physical AI could reduce resource needs for some tasks and expand services. Fiscal institutions must connect that capacity with sustainable public services and transition support. Long-term business returns also depend on solvent customers, reliable public services, and people able to participate.
Robots may work without wages. An automation business model cannot assume society needs no income.
If wages become a smaller share of income, would you favor changes to capital and consumption taxation or a specific automation tax? Which cost concerns you most?
Sources current to October 20, 2026. This essay discusses policy design, not an existing uniform robot tax implemented by any country.