Charging levies on AI 'tokens'—a usage tax applied at the point of inference—could give governments a practical way to tax AI-generated value and blunt automation-driven revenue and wage shocks; the approach hinges on robust cryptographic receipts, norm-based rates, white-box audits, and difficult international cooperation.

Main Finding

Token taxes—usage-based surcharges levied on model inference “tokens” at the point of sale—are a practical, enforceable policy instrument for mitigating the major economic risks of AGI (shrinking tax bases, falling living standards, and citizen disempowerment). Compared with robot- or FLOP-based taxes, token taxes better capture where AI-generated value is realized and can be enforced using existing compute-governance and commercial billing infrastructure. A staged audit pipeline (black-box token verification → norm-based tax rates → white-box audits) combined with agent-based modeling of macro and distributional effects provides a feasible path to design and evaluate such taxes.

Key Points

• Motivation

  • AGI-related automation can drastically reduce labor income and erode government tax bases; the problem is qualitatively larger than first-industrial-revolution wage stagnation.
  • Existing AI-safety work has emphasized capability risks; comparatively little rigorous work targets economic mitigation policy.

• What token taxes are

  • Usage-based surcharge applied at the point of sale for model inference (i.e., charged per token or per inference request).
  • Intended to tax value creation where models are used (recipient-facing), not just where models or hardware are hosted.

• Advantages relative to alternatives

  • Enforceability: can leverage existing commercial and compute-governance infrastructure (API billing, cloud metering, hardware telemetry, attestation).
  • Value capture: ties taxation to user-facing consumption rather than to physical location of compute or model hosting, limiting simple geographic tax arbitrage.
  • Simpler to align with retail/payment systems than taxing intermediate inputs (like FLOPs or robots).

• Enforcement architecture (staged audit pipeline)

• Black-box token verification: cryptographic/ledgered tokens or receipts accompany inference responses to prove taxable consumption without full model inspection.
• Norm-based tax rates: standardized tax schedules based on observable usage characteristics (e.g., tokens per task, agentic capability categories) to reduce gaming and simplify compliance.

• White-box audits: deeper inspections (model internals, logs, provenance) triggered by anomalies or high-value activity to detect evasion and recalibrate norms.

• Modeling & evaluation needs

  • The paper emphasizes constructing agent-based models (ABMs) to simulate micro-to-macro dynamics: firm adoption, labor reallocation, price pass-through, tax incidence, cross-border flows, and long-run growth impacts.
  • ABMs needed because standard representative-agent or DSGE models may miss heterogeneity, network effects, and path dependence.

• Alternatives & political economy

  • FLOP taxes: tax raw compute (FLOPs) but face measurement, enforceability, and leakage challenges; they tax inputs rather than where value is realized.
  • Risk of “AI superpower” veto: dominant states or firms could block or undermine coordination. The paper discusses mechanisms to deter vetoes (trade measures, hardware export controls, multilateral agreements, leveraging cloud provider dependence), but highlights high political-economy complexity.

Data & Methods

• Nature of the work

  • Policy-design and conceptual-architecture paper rather than empirical analysis. No original microdata or econometric estimates are presented.
  • Methods are primarily normative and engineering-policy: specification of tax base, enforcement pipeline, and a research agenda for simulation.

• Enforcement technical components described

  • Black-box verification mechanisms: issuing tamper-evident consumption tokens or receipts tied to API calls/inference events, potentially leveraging cryptographic signatures or ledgers.
  • Norm-based taxation: deriving standardized tax schedules from observable proxies of capability/impact (e.g., token intensity, response latency, declared agentic features).
  • White-box auditing procedures: legal and technical protocols for inspecting model weights, training provenance, and inference logs where necessary.

• Modeling proposal

  • Calls for building agent-based models to quantify:
    • Incidence of token taxes on wages, profits, prices.
    • Dynamics of unemployment, retraining, and reallocation.
    • International arbitrage and firm relocation incentives.
    • Welfare trade-offs across income groups and regions.
  • Comparative modeling of token taxes vs FLOP taxes and other interventions (universal basic income, wage subsidies, capital taxes).

• Limitations acknowledged

  • No empirical calibration provided; effectiveness depends on assumptions about evasion costs, measurement integrity, and international coordination.
  • Technical feasibility of robust token verification and resistance to spoofing needs demonstration.
  • Political feasibility—especially preventing veto by concentrated AI powers—is uncertain and under-specified.

Implications for AI Economics

• Public finance & tax policy

  • Token taxes offer a new tax base tightly linked to digital value creation by AI, potentially restoring revenue lost to automation.
  • The choice of tax base affects incidence: tokens tied to consumption likely shift burden toward AI service buyers/end-consumers and AI capital owners differently than FLOP or corporate taxes.

• Labor markets & distribution

  • Depending on incidence and price pass-through, token taxes could slow displacement by increasing the effective cost of automation, buying time for retraining and redistribution.
  • ABMs are essential to identify distributional outcomes—who bears the tax burden, how wages evolve, and whether revenues can fund safety nets.

• Innovation & firm behavior

  • Careful rate-setting needed to avoid stifling productive innovation or driving compute/modeling offshore.
  • Token taxes incentivize more efficient model designs (fewer tokens per task) and may shift competition toward lightweight models or on-device solutions.

• International coordination & governance

  • Token taxes reduce some arbitrage but do not eliminate cross-border avoidance; multilateral agreements and trade/regulatory levers will be crucial.
  • Preventing concentrated actors (states or firms) from vetoing enforcement requires building broad coalitions, linking compliance to hardware and cloud access, and embedding rules in trade/standards regimes.

• Research & policy next steps (operational)

  • Build and calibrate ABMs with agent heterogeneity (firms, workers, governments, consumers) to test tax rates, compliance costs, and secondary effects.
  • Prototype technical implementations of token verification (proof-of-query receipts, cryptographic attestation) and red-team them for spoofing/evasion vectors.
  • Comparative analysis: token taxes vs FLOP taxes, payroll/corporate taxes, subsidies, and non-tax measures (universal basic income, retraining).
  • Institutional design: legal frameworks for audits, privacy-preserving verification standards, and mechanisms to reduce the risk of geopolitical veto.

Concluding note: Token taxes are a promising, operationally grounded lever for addressing AGI-driven economic disruption, but they require technical validation, detailed economic modeling (especially via ABMs), and robust international governance to be effective and politically feasible.