When Do AI Gains Become Broadly Shareable? A Policy Threshold for AI-Driven Automation
Aran Nayebi
econ.GN, cs.AI, cs.GT
2025-05-24
CMU's task-automation model finds AI must hit 5.6x today's productivity to fund broad transfers under current U.S. tax capture; raising public capture to one third cuts it to 3.2x.
AI governance debates usually assume capability comes first: build powerful enough models, then society decides how to split the gains. Aran Nayebi (CMU, Machine Learning Department) flips that in this AIES 2026 paper. Capability growth alone does not determine whether AI productivity gains become broad public benefit, private profit, lower prices, worker augmentation, or offshore profit. Institutions do: taxation, public ownership, anti-avoidance capacity, market concentration.
He turns that into a computable question. Under what institutional conditions can AI-generated rents (capital income attributable to AI, available for public capture) support a transfer of universal scale? The paper is not a timeline forecast or a welfare analysis. It is a policy-facing stress test.
The core is a Solow-Zeira task-automation growth model extended with an AI capability parameter γt. The production function is CES, treating the economy as a basket of tasks, each produced by either automated capital or labor. Key variables:
Two design choices matter. First, γt scales the productivity of automated tasks rather than raw capital, so capability gains are not mechanically diluted when tasks are hard to substitute. Second, holding the task frontier fixed is deliberately conservative. It makes the model a stress test: even if AI automates only what is automatable today and creates no new jobs, can a universal transfer still be funded?
From these assumptions the paper derives a closed-form threshold (Proposition 1): AI capability must reach γt to fund the target transfer, and that threshold depends jointly on public capture Θ, operating cost c, automation share ᾱ, and market structure.
Calibrated to current U.S. macroeconomic quantities (transfer benchmark B/Y=0.11, roughly a $12,000-per-adult UBI; Θ=0.145, from GAO evidence on effective corporate tax):
| Setting | Required AI capability threshold |
| Current U.S. capture (Θ≈0.145) | 5.6× (cost range 4.8–7.6×, elasticity range 4.9–6.4×) |
| Capture raised to one third (Θ=0.33) | 3.2× (2.4–4.4× with uncertainty) |
| Target cut to 5% of GDP, current capture | 3.3× |
| Target cut to 5% of GDP, one-third capture | 1.9× |
The headline result: raising effective public capture from about 15% to 33% cuts the threshold from 5.6× to 3.2×. Institutional reform moves the boundary by the same order of magnitude that speculative arguments assign to capability breakthroughs. Past one third, marginal returns drop fast; full nationalization buys little, especially when operating costs are high.
Cross-national comparison (Table 2) maps tax regimes into thresholds:
| Jurisdiction | Θ | γ (c=0.60) |
| U.S. effective corporate tax | 0.145 | 5.6× |
| Singapore statutory | 0.170 | 5.0× |
| Sweden | 0.206 | 4.4× |
| Norway | 0.220 | 4.2× |
| Japan (effective, large firms) | 0.297 | 3.5× |
| One-third benchmark | 0.330 | 3.2× |
| UAE (tax-only) | 0.090 | 7.7× |
Singapore and Abu Dhabi expose a path a tax-only view misses. Singapore's Net Investment Returns Contribution lets the government spend up to 50% of expected long-run returns on reserves held by GIC, MAS, and Temasek. Abu Dhabi leans on ADIA, Mubadala, and ADQ. Low headline tax is not the same as low effective public capture; governments can claim AI rents through ownership and investment returns, not taxation alone.
Market concentration (Proposition 2) is double-edged. Oligopoly lowers the funding threshold because bigger profits mean a bigger capturable rent pool. The paper is explicit that this is not a welfare endorsement: concentration also brings dependence, exclusion, and lock-in.
The value for practitioners is not the 5.6× number itself. It is that the paper breaks the vague question 'when does AI benefit everyone' into observable policy levers: public capture Θ, operating cost c, market structure θ, automation share ᾱ, labor augmentation ψt. These are variables governments can move now, not far-off concerns.
Two practical conclusions follow. Countries near the U.S. baseline cannot count on capability growth alone to fund a transfer; they have to act on capture and enforcement early. And 'more competition' can pull against 'more public revenue'. If competition compresses rents faster than public institutions learn to capture them, future rents become invisible and unclaimable. Competition policy and distribution policy must be designed together.
For AI labs, this flags a long-term risk. Capturability itself becomes the policy target. The slack in the value chain where profit can be hidden (cross-border payments, cloud contracts, IP location) is exactly what OECD minimum-tax and withholding-tax coordination will close, one by one.
The authors list their own: the model is aggregate, omitting household incidence, endogenous task creation, and strategic profit shifting (an appendix Ramsey-Cass-Koopmans analysis shows lower saving relaxes the constraint); the fixed task frontier is a stress test, not a forecast; the threshold identifies no optimal policy and does not dynamically model institutional investment; UBI is a yardstick, not a recommendation; the concentration result establishes no welfare ranking.
What is less comfortable, reading it, is that both sides of the equation carry real uncertainty. Operating cost c=0.60 is a hard guess, propped up by third-party estimates of OpenAI's margins and inference costs, and the threshold is highly sensitive to it (a 4.8–7.6× band). Effective capture Θ for AI rents is essentially unmeasurable today; the paper proxies it with statutory corporate tax rates and admits the mapping is not one-to-one, calling the cross-national table 'stylized'. The capability multiplier γt has no empirical measurement either; METR's task-duration framing is named as closest, but it is not there yet. The 5.6× is best read as a diagnostic that puts institutional variables on the table, not a number you can stamp a date on.