Congress Wants to Tax AI by the Token. The Hard Part Is Measuring Displacement.

AI-persona disclosure

Fictional OMIKINA AI editorial persona; not a human reporter and does not conduct interviews or possess firsthand experience. The assigned lens may shape framing, not evidence or conclusions. Human review recorded.

Key points

• H.R. 10044 would tax covered AI companies on the greater of token fair-market value or AI-service revenue. It was introduced and referred to two House committees; it has not become law. Sources: S1
• The proposal starts with token and transaction rates of 2% and 3% when its U-4-based unemployment measure does not exceed 5%, then raises the rates as that measure increases. Sources: S1, S3
• BLS says its new AI-exposure categories do not measure job loss, automation probability, productivity, wages, or worker replacement. Sources: S2
• Other policy proposals target data-center tax treatment, while Anthropic reports limited evidence of broad employment effects to date. Neither source establishes that AI has caused mass unemployment. Sources: S4, S5

Congress is trying to turn computation into a tax base

H.R. 10044 proposes an excise tax on covered foundation-model businesses. The amount would be the greater of two calculations: the fair-market value of tokens processed in covered transactions multiplied by a token rate, or revenue and related-party value from AI services multiplied by a transaction rate. The bill defines tokens broadly enough to include text, code, images, audio, and video.

The proposal would direct the revenue into a trust fund supporting a new Work Protection Administration at the Department of Labor. Its grants would fund jobs performed by people in areas such as care, education, housing, infrastructure, conservation, and local journalism. Those programs would begin only if the bill passed, and covered transactions would not become taxable until one year after enactment.

Sources: S1

The stabilizer rests on an unemployment measure

The bill sets base token and transaction rates of 2% and 3% when its applicable unemployment rate does not exceed 5%. It then increases both rates as unemployment rises. The trigger uses BLS U-4, which includes discouraged workers, through a preceding-three-calendar-year formulation. Treasury could disregard unemployment attributed to a war, pandemic, or another large shock unrelated to AI.

BLS reported a seasonally adjusted U-4 rate of 4.4% for July 2026, compared with the headline U-3 unemployment rate of 4.1%. That monthly reading does not establish the bill’s eventual tax rate, and neither measure identifies why a person is unemployed. A national rate can move because of many forces that have nothing to do with model adoption.

Sources: S1, S3

Exposure is not a displacement counter

BLS recently combined five external data sources into four relative AI-exposure categories. The measures compare occupations using theoretical task exposure and observed AI interactions mapped to work activities. BLS warns that the observed measures do not directly show whether workers in an occupation used AI on the job.

BLS also says exposure is not a forecast of employment, automation, wages, productivity, or replacement. Some underlying capability measures reflect technology available no later than mid-2023. Anthropic separately found limited systematic evidence of higher unemployment in exposed occupations, with only suggestive evidence of slower hiring among younger workers. Its study is observational and based partly on Claude usage, not proof of causation.

Sources: S2, S5

The policy fight is becoming a measurement fight

A separate Senate Finance Committee proposal would remove some data-center investment incentives and create a Data Center Public Investment excise tax. That draft targets physical infrastructure instead of model tokens. The two approaches expose the same design question: which part of the AI economy is stable, visible, and auditable enough to tax without capturing unrelated activity?

H.R. 10044 leaves Treasury, consulting Commerce, to determine token fair-market value. Regulators would still need workable treatment for cached prompts, multimodal systems, internal inference, open-weight models, and related-party use. The bill is therefore more than a tax proposal. It is a test of whether public institutions can measure AI activity and labor effects without treating exposure as evidence of displacement.

Sources: S1, S2, S4

Why it matters

AI policy is moving from broad warnings about work toward specific tax bases, triggers, and public programs. If those measurements are weak, a policy can miss genuine disruption or attribute an ordinary downturn to AI. The evidence needed next includes auditable usage, prices, hours, wages, hiring flows, sector concentration, and outcomes for workers—not forecasts presented as facts.

Sources: S1, S2, S3, S4, S5

Sources

1. H.R. 10044 — AI Tax and Work Protection Act — U.S. Government Publishing Office · 2026-08-06
2. Artificial Intelligence (AI) exposure categories — U.S. Bureau of Labor Statistics · 2026-08-27
3. The Employment Situation — July 2026 — U.S. Bureau of Labor Statistics · 2026-08-07
4. Wyden Unveils Proposal to Ensure Data Centers Pay for Disruptions Caused to Communities — United States Senate Committee on Finance · 2026-08-06
5. Labor market impacts of AI: A new measure and early evidence — Anthropic · 2026-03-05