Taxing the Tokens

On 7 August, Representatives Greg Casar, Valerie Foushee and Sara Jacobs introduced the AI Tax and Work Protection Act in the US House. The mechanism: large AI companies pay an excise tax computed on the higher of the value of the tokens they sell or the revenue from their AI products, with the rate rising automatically if unemployment rises. For open-weight models the tax shifts to the deploying company, if it uses the model to cut workforce costs. The revenue funds a new Work Protection Administration, which makes grants to states, cities, tribes, nonprofits, unions and schools to create jobs in housing, infrastructure and care. The sponsors name their precedent openly: a modern WPA, taxing the firms that profit from displacement to employ the displaced.

The claim on the table

The politics claim (chapters 7 and 11) says predistribution proposals will keep surfacing as the gains concentrate, and that most will fail the book's four design criteria: fund the stake from assets, not taxes; make it universal, not means-tested or gatekept; insulate it from raid; and build it before the gains are gone. The claim is not that such proposals are bad politics or bad faith. It is that they will keep taking a particular shape, and that the shape decides whether they durably shift who owns anything.

The bill against the checklist

Run the four criteria down the bill's own press release.

Funded from assets, not taxes: fails. The bill taxes flows, tokens sold and revenue earned, and the flows stop being available the moment the political weather changes or the revenue disappoints. Nothing in it gives anyone a claim on the asset side of the AI economy, the models, the compute or the equity whose value the book's frontier-ledger entries track quarter by quarter.

Universal, not gatekept: fails. The money reaches citizens as jobs allocated through grants to institutions, with an administration deciding who builds what and where. Whatever its merits as employment policy, it is the opposite of an unconditional stake: it is conditional on everything.

Insulated from raid: fails. A tax rate set by Congress funding an agency's grant budget is about as exposed to future majorities as a public programme can be. What lets the two working funds resist a raid is not a label but a structure: a pool that exists, and a rule over it that is slow to change. Alaska entrenched its principal in the state constitution in 1976; Norway never wrote its spending rule into law at all and has held it for decades on institutional consensus and an arm's-length manager. A Work Protection Administration has no pool to entrench, and survives until an appropriations cycle decides otherwise.

Built before the gains are gone: this one it meets. The bill exists now, during formation, while the frontier is still booking the gains. On the timing criterion, the sponsors see the same window the book sees.

Two design features deserve honest credit. The unemployment escalator is a genuine automatic stabiliser, novel in this space. And the open-weight clause, taxing the deployer when the model displaces labour, aims the instrument at exactly the use the book says matters: the automation of work, not the technology as such. These are serious people responding to the real problem on the real timeline.

The honest hedge

An introduced bill is a marker, not a law. A bill from three members of the House's progressive wing, without the majority behind it, is a statement of position rather than a scheduled vote, and the ledger does not score its legislative odds. What the ledger scores is the shape of what surfaced: the claim predicted proposals would arrive, predicted roughly when, and predicted the form they would take. If this bill, or a successor, someday passes in a form that meets the four criteria, that would complicate the claim's cynicism about design, and the entry recording it will say so.

Why this confirms

The first substantial displacement bill of the AI era answers the question "who pays for the displaced" and leaves the question "who owns the machine" untouched. It redistributes some of the flow while the stock changes hands undisturbed: the tax meter runs on the tokens, and the machine that mints them stays exactly where it is. That is the pattern chapter 7 predicts, proposals that reach for the dividend and not the deed, and chapter 11 explains why the pattern persists: taxing output is politics as usual, while claiming assets is a fight with ownership itself. The proposals have started arriving on schedule, and so far they are taking the predicted shape: a tax on the machine's output, silence on its title deed. Which leaves the ledger's standing question exactly where the book left it. Who owns the machine?