There are two ways to close the gap: attach a cost to machine work, or reward employing people. The tax side is the larger and louder half — every serious proposal answers the same four questions in a different order, and lands in one of eight categories depending on the answers. The incentive side is smaller, quieter, and the only part with bipartisan sponsorship today. Read the sequence first, then the comparison, then the incentive side.
Read left to right: the thing being measured, the moment tax attaches, who writes the check, and where the money is earmarked.
Who remits is not who bears it. Incidence is the hardest fight in every one of these bills.
Grouped by what they tax rather than who proposed them. The three ratings are the axes economists argue over most: whether the base tracks actual usage, how hard it is to audit, and how much it risks discouraging investment.
RATINGS SYNTHESISED FROM PUBLISHED COMPARISONS INCL. ABBOTT & BOGENSCHNEIDER (2018), BOGENSCHNEIDER IN TAX NOTES (NOV 2025), THE TOKEN-TAX POSITION PAPER (ARXIV 2603.04555) AND KORINEK & LOCKWOOD FOR BROOKINGS (JAN 2026). SEE RESEARCH.
Every mechanism above works by attaching a cost to machine work. The mirror approach leaves AI alone and rewards employing people instead — a smaller, quieter set of proposals, and the only part of this debate with bipartisan sponsorship on the board today.
The asymmetry is the whole argument: a company that cuts an employee stops paying payroll tax, while the system doing that work pays none. H.R. 10044 attacks it from both ends, with rates that escalate automatically when national unemployment passes five percent.
Lowering employer payroll tax does the same work without touching AI at all — and it is the version economists across the spectrum find easiest to defend.
CASAR OP-ED, MAY 2026 ↗The AI Workforce Training Act — Reps. Gottheimer (D-NJ) and Lawler (R-NY) — would let companies claim a credit worth 30% of qualified expenses for teaching employees to use, manage, and build AI systems, capped at $2,500 per employee a year.
Accredited courses, certificate programs, and workshops all count. It is the rare measure here with a Republican co-sponsor.
FEDSCOOP, FEB 2026 ↗The Work Opportunity Tax Credit — the federal government's standing subsidy for hiring people who face barriers to employment — has been in a legislative pause since its authorization expired on 31 December 2025.
Without new legislation, employers generally cannot claim it for anyone who started in 2026. Renewing it is the fastest available move on this side of the ledger, and it needs no new tax to work.
IRS — WOTC ↗Tax practitioners advising companies read current federal law as having changed the economics of workforce planning in favour of automation, and describe a new round of state competition in which jurisdictions offer AI and automation investment credits rather than labor-based ones.
So the incentive side is not starting from neutral. Any credit for employing people is being written on top of a code that already rewards the opposite choice — which is why most sponsors pair a credit like these with one of the eight categories above.
ELLIOTT DAVIS, DEC 2025 ↗Not every proposal arrives as a bill or a working paper. These are the ideas being argued in essays, interviews, and public letters — no sponsor, no bill number, no peer review. They are here because they are shaping the debate, and because the objections to them are the sharpest part of it.
Tax the machines that replace workers and use the revenue for retraining and a stronger safety net. Asked how it would work, Gates put a number on it: half of a company’s token-tax revenue paid to government, held to help people who lose their job to AI. He concedes it is not economically efficient, and argues that with innovation accelerating we can afford some inefficiency as the price of keeping people employed.
MIT TECHNOLOGY REVIEW ↗Not a tax at all: hold certain work for people by agreement — caregivers, mental health workers, some medical professionals — with the list varying country to country. The idea came from watching the caregivers who looked after his father, work he calls irreplaceably human. Axios names the practical problems plainly: which jobs qualify, who decides, what stops companies cheating, and how a country protects those jobs if competitors abroad automate them.
AXIOS ↗Etzioni backs taxing robots that replace workers but not tokens: “I think that taxing tokens is like taxing keystrokes on a typewriter. It doesn’t make sense. It measures effort, not displacement of workers.” He also warns that a tax designed to slow adoption is a hard thing to run in a competitive field: you tax what you want to happen more slowly, and then you have to ask who overtakes you.
ABC NEWS ↗A per-token levy is collectable at a metering boundary: an API bills you, so a government can tax the bill. Tokens produced by an open-weight model on hardware you own cross no such boundary — no invoice, no counterparty, no meter — so the tax may simply push work onto local models. This is the argument any usage-based proposal has to answer, and it is why several sponsors pair a levy with the reporting mandates on the Proposed Solutions page.
ANALYSIS OF THE ESSAY ↗LISTED BECAUSE THEY ARE BEING ARGUED IN PUBLIC, NOT BECAUSE THEY HAVE BEEN INTRODUCED, COSTED, OR ENDORSED. NOTHING HERE HAS A BILL NUMBER.