The richest worker of the AI age will never file a W-2. When production shifts from payrolls to platforms, a tax code built to skim labor income loses its anchor. The current debate asks whether underemployment will arrive. The more useful question is whether the revenue system will still function when it does. We do not have to assume science fiction to see the fault lines. AI centralizes capability, accelerates cycles, and hides risk in code. That combination defeats tax regimes that rely on slow, human rhythms. A new code is necessary, but not in the way most proposals imagine.
The economics are blunt. AI-driven wealth is capital intensive. Training large models demands scarce compute, proprietary data, and expensive engineering. That tilts returns toward owners of chips, clouds, and platforms. Payroll taxes then shrink, even as output rises. UMA and others point out that this capital barrier fences out small entrepreneurs and concentrates gains in a few balance sheets. We have seen mechanization move returns from labor to capital before, but the AI curve is steeper. Algorithms scale with near-zero marginal labor. Underemployment is not a moral panic. It is a balance sheet reality when one engineer with an agent farm replaces a floor of analysts.
A tax base should target where economic density forms. In the last century, that was the factory and the office. In this one, it is the model, the data center, and the API gateway. A payroll-centric system will weaken as the nexus of value moves. Waiting for mass dislocation to confirm the theory would be policymaking by rearview mirror.
Calls for an AI wealth tax are back. Historically, wealth levies have followed shocks and wars, and the instinct feels right in a moment of fast concentration. But the administrative track record is poor. Valuation complexity, capital flight, and avoidance pushed many economies to drop them. The Hoover critique is not wrong on logistics, even if it underweights distributional concerns. Assets in an AI economy are more intangible and mobile than ever. You cannot inventory a model weight file the way you count factories. Chasing balance sheets risks taxing the shadow, not the object.
Target the choke points instead. Tax the gateways that cannot move offshore overnight: high-end compute consumption, hyperscale cloud invoices, training runs above specified thresholds, and platform gross receipts from AI APIs. Think like an engineer. Reinforce the stress points you can measure and audit, not the diffuse outputs. Build exemptions and credits for safety, open science, and small entrepreneurs to avoid entrenching incumbents. If you must tax wealth, do it as a time-limited windfall levy on clearly abnormal profits defined by rate-of-change triggers, with hard sunsets. It is easier to measure surges than levels.
Markets are not ready for agent herding. The Roosevelt Institute and others have flagged a simple game theory problem. When many AI agents train on the same data and optimize on similar signals, they converge on correlated strategies. That reduces diversity in decision rules and sets up synchronized errors. The result is higher, fatter-tail volatility, flash crashes, and sudden regime shifts when models adapt in unison. Add manipulation risk from autonomous agents that can spoof, flood, or front-run, and you have a market machine that can amplify noise into systemic moves.
Tax receipts ride these waves. Capital gains taxes, corporate taxes, and even sales taxes have become procyclical, tied to asset prices and sentiment. In an AI market, the cycle will run faster and deeper. A robust code needs automatic stabilizers funded by the AI rents it skims in the boom. Think of a reinsurance pool for financial stability, financed by levies on platforms and compute hubs, that auto-deploys when volatility or drawdowns cross defined thresholds. Design the fiscal circuit breaker before the first multi-agent flash crash triggers a revenue drought.
Tax authorities will automate. That is rational. But AI in tax administration introduces legal and political fragility. State policy analysts already warn that algorithmic tools can replicate or worsen existing biases, especially against small businesses and marginalized groups. Expanded analytics raise significant privacy concerns through over-collection and broader attack surfaces. Worse, opacity undermines due process. If a model flags you for audit or denies a credit, how do you confront evidence you cannot see or interpret?
A tax system must be contestable to be legitimate. Any AI deployed for compliance or assessment should be licensed, logged, and subject to independent audit. Decisions affecting obligations should come with plain-language explanations and a right to human review. Adopt data minimization by default. Use provable privacy tools and red-team models for disparate impact before production. Build rollback plans. The point is not to freeze the administrator in analog time. It is to prevent the tool from delegitimizing the very state capacity it is meant to enhance.
Taxation rests on consent as much as code. If people believe their financial lives are ingestible, traceable, and leaky, they will rearrange behavior. Capital and talent seek opacity when trust erodes, and the tax base follows. UMA’s warning about privacy violations and bias is not a compliance memo. It is a forecast for revenue.
Guardrails are industrial hygiene. Segregate taxpayer data from training corpora. Mandate bonded custodianship for sensitive financial data used in tax analytics. Require differential privacy or comparable protections for any statistical use. Penalize unauthorized access on a strict liability basis. The fines must be large enough to register on the P and L of a hyperscaler. Skimping on this turns every breach into a tax avoidance seminar.
If income migrates to AI platforms, the code should meet it there. A compute excise, triggered above a high threshold of training or inference usage, is measurable and enforceable. Cloud providers already meter GPU hours and bandwidth. That is an audit trail. A small per-unit levy on compute for frontier-scale workloads, paired with generous de minimis rules, would capture industrial rents without strangling startups. Alternatively, tax platform gross receipts from AI APIs, indexed to volume and latency classes, with credits for open-source contributions and safety evaluations.
This is less romantic than a robot tax or a sweeping wealth levy, and more durable. Bottlenecks are the natural points of enforcement in any system. Chips, cloud, and platforms are the valves in this economy. Price the valves. Use the proceeds to lower payroll taxes on human labor to slow underemployment’s spread and to fund lifelong learning. Tie credits to firm behavior that complements, not replaces, human workers.
Systems that benefit from stress adapt through options, not plans. The labor market will need options. Replace blunt subsidies with contingent instruments. Offer time-bound hiring credits for firms that add net human roles in AI-augmented functions. Introduce a negative payroll tax for low-to-middle wage jobs in sectors at high risk of displacement, financed by the compute or platform levy. Create a standing prize pool for tools that improve model safety, auditability, and energy efficiency, rather than pre-picking vendors.
Automaticity matters. Index supplemental unemployment insurance to local measures of displacement derived from tax filings and job postings, with hard caps and expirations. Pre-fund retraining accounts that unlock when a worker’s prior occupation sees defined declines in postings. These are not acts of charity. They are system stabilizers that turn volatility into adaptation instead of shock.
AI concentrates wealth and power. That is not ideology. It is the predictable result of scale economies in compute and data. The AI Safety Atlas frames this bluntly. If we ignore it, politics will eventually reach for blunt force. Better to shape the channel now. Enforce interoperability where natural monopolies form. Use structural separation when a platform controls both the rails and the trains. Mandate model and dataset escrow for critical systems to ensure continuity if a provider fails.
On tax, resist nostalgia and moral theater. Wealth taxes reappear after crises because the distribution looks intolerable. They also underperform because they chase mobile, complex assets. A more credible plan is a time-limited excess profits levy triggered by clear parameters, combined with ongoing taxes at the compute and platform choke points. Pair that with strict privacy law for tax analytics, audited AI use in administration, and market stability reserves funded by the very agents that heighten volatility. That is not punitive. It is basic resilience engineering for a system whose load profile just changed.
The false comfort is to wait for underemployment data and then react. By then, the base will have shifted, the politics will be inflamed, and the tools will be entrenched. The contrarian position is to tax the bottlenecks, insure against correlated failure, preserve due process in automated enforcement, and buy options for labor. Build a code that gets stronger when the models get sharper.
