Harvard study warns AI could shrink tax revenue and widen inequality

Researchers from Harvard University and the Brookings Institution published the study "How AI Development Affects Fiscal Policy," which projects the effects of automation on public budgets, employment and inequality through 2060. The central conclusion, reported by Bloomberg and summarized by Russian outlet CNews, is that governments risk facing a drop in tax revenue at the same time they need to increase spending on support and retraining for displaced workers. The authors argue that countries should begin reforming fiscal policy as early as 2026 to better absorb the structural shocks to come.

The point that matters for those who build software: AI's fiscal effect doesn't depend only on how much it increases national income, but on who captures that income and how governments respond. And the technology sector is at the center of this dispute.

The study's foundation and the four scenarios

The projections are based on the long-term estimate from the U.S. Congressional Budget Office (CBO). In the baseline scenario, U.S. public debt rises from 101% of GDP in 2026 to 175% of GDP in 2056, already factoring in an annual productivity gain of 0.1 percentage point through 2036 thanks to AI. On top of this baseline, the researchers built four additional scenarios:

• Scenario 1 (high, broadly distributed productivity): productivity growth of 0.5 pp per year. The economy expands by about 7 pp, tax revenue grows, and debt reaches 136% of GDP by 2060, 39 pp below the baseline. Part of the relief comes from social spending that is indexed to inflation, not to wages.
• Scenario 2 (gains concentrated at the top): same 0.5 pp productivity, but the benefits flow to high-income groups. Inequality rises, but federal revenue increases because of the U.S. progressive tax system (average rate of ~21% overall and ~26% for the top quintile). GDP 5% higher after 10 years and 16% after 30 years; debt at 126% of GDP.
• Scenario 3 (significant unemployment): about 3 million workers displaced (2% of the roughly 170 million employed). Many don't find comparable new jobs. To contain social tension, the government expands unemployment insurance, creates wage insurance and expands retraining programs, which erodes part of the fiscal gain.
• Scenario 4 (doubled automation): productivity of 1 pp per year and double the AI penetration. Six percent of workers lose their jobs, and income structurally shifts from labor to capital. GDP surges (9% higher by 2040, 21% by 2050, 35% by 2060), but the budget gains little: in the U.S., capital income is taxed about 40% more lightly than labor income.

The paradox of the fourth scenario is the study's strongest message: even with much higher productivity, debt still rises to 126% of GDP because the tax base has slipped toward capital, which is taxed less. As a response, the authors cite options ranging from higher taxes on capital, wealth and consumption to universal basic income and public-sector job creation.

The shift from labor to capital is not new

The decline in labor's share of GDP in developed economies has been observed since 1980, according to analysts cited in the report. The study formalizes a long-standing problem of tax systems: over the long run, capital taxes tend to be pushed down to avoid discouraging investment, which shifts the burden increasingly onto labor, precisely the factor that AI threatens to reduce.

A concrete data point reinforces the thesis: companies in the U.S. are already reporting in quarterly earnings higher margins from AI use alongside simultaneous staff cuts. If payroll shrinks while margins rise, the tax revenue that used to come from wages disappears, while the revenue that would come from profits is taxed more lightly.

What changes for Brazilian developers

The study is about the U.S., but the risks are "absolutely relevant" for other countries, and there's one detail that hits the IT sector directly. According to a survey by Sberbank's Center for Macroeconomic Research cited in the report, 48% of the Russian economy has high exposure to AI, with IT leading the way (potential to cover nearly 60% of the labor shortfall), followed by the financial sector (just over 50%) and research, public administration and retail (about 40% each). The analysis mapped 18,000 tasks across 923 occupations.

The reading for Brazil follows the same structural logic. Here, state and municipal tax revenue depends heavily on taxes tied to income and local economic activity, while the value generated by AI platforms tends to concentrate among a few owners of technology capital, often outside the country. The risk vector flagged by analysts in other markets, the concentration of the tax base at the top and the transfer of resources from the regions to the center, is an understandable warning for Brazil's federative pact (the constitutional arrangement for sharing tax revenue among federal, state and municipal governments), which depends on transfers and payroll-based taxes.

There's also the tax-incentive angle. The report notes that Russian legislation began applying a 2.0x multiplier for AI investments, encouraging deductions and legal minimization of corporate taxes, exactly the "accelerated enrichment of technology capital owners" effect the study warns about. It's worth comparing this to the design of innovation incentives in Brazil (such as the Lei do Bem, Brazil's R&D tax incentive law, and software taxation regimes): anyone designing products and pricing services around automation needs to understand that these incentives may change shape once the fiscal debate matures.

What remains open

The authors themselves treat the topic as one of the central issues of economic policy this decade, and the proposed answers conflict with each other: raising capital taxes reduces investment, taxing consumption (through something equivalent to VAT or a sales tax) weighs on those with less, and universal basic income requires financing that remains unresolved. For developers, three fronts stay on the radar: how Brazil's tax reform, currently being implemented, will treat digital services and automation; whether there will be pressure to tax more profit and less payroll; and how AI incentives will be calibrated so they don't simply become a tax cut for those who already concentrate capital. None of this is settled: the authors treat the topic as one of the central issues of economic policy of our time, without pointing to a specific timeline for the outcome.