The scenario hands its government two bills, and both lose
Citrini Research's The 2028 Global Intelligence Crisis is a macro memo written from June 2028, published in February 2026, and labeled by its authors "a scenario and not a prediction." Most readers took away the market numbers: 10.2% unemployment, a 38% drawdown. The fiscal chapter is more useful. The authors give their fictional administration two bills. The Transition Economy Act funds "direct transfers to displaced workers" through deficit spending and "a proposed tax on AI inference compute." The Shared AI Prosperity Act "would establish a public claim on the returns of the intelligence infrastructure itself," something "between a sovereign wealth fund and a royalty on AI-generated output."
In the story, the two bills are debated while federal receipts run "12% below CBO baseline projections" and the right and left split along familiar lines. Both arrive after the damage. That is the point of a scenario: the policy that would have worked is visible only in hindsight. So treat the two acts as a design exercise. Score them the way a finance ministry would, on base, payer, trigger and erosion, and see which one would collect if a legislature wrote it this year.
A four-part scorecard for any AI tax
Every tax on autonomous work has to answer four questions. Base: what is measured? A base that already exists in the accounts, such as revenue, is cheaper to administer than one that has to be invented. Payer: who writes the check? If it is not the party that captured the value, the tax is passed along or avoided. Trigger: when does the rate move? A fixed rate in a fast-moving transition is wrong in both directions. Erosion: does the base shrink as the technology improves, or grow with it?
Add the two gates from the open-source shift of 2026. Does the design collect the same amount whether the model was open or closed? Does it collect the same whether an application was bought or cloned? Those gates matter because this year Meta shipped Muse Spark as "proprietary only" while REA let any agent read a compiled binary and "build a version for your project." A base that depends on how the code was obtained will be rewritten by the next licensing decision.
The Transition Economy Act: a tax that erodes under its own feet
The inference compute tax scores well on exactly one line. It is easy to measure. Compute is metered by the providers already, and a tax authority could read the meter. Everything else fails. Payer: the lab or cloud provider pays, not the insurer that emptied a claims floor. The cost passes through to token prices, which the deployer absorbs or avoids by running open weights on its own hardware, where no provider meters anything.
Erosion is the fatal line. Meta says Muse Spark reaches its reasoning capabilities with roughly ten times less compute than Llama 4 Maverick. Take that at face value and a compute tax collects a tenth as much for the same work, across one model generation. The labs compete on exactly this number. A tax on compute is a tax whose base every engineer in the industry is paid to shrink. In Citrini's own scenario, the right objects that taxing compute hands the lead to China. The more precise objection is that the base vanishes without anyone's help.
Trigger: the act has none. It is funded by deficit spending plus the compute tax, which means it is sized by politics, not by displacement. Citadel Securities' rebuttal to the scenario argued that if compute costs more than human labor, "substitution will not occur." Read that as a warning about the act. A compute tax raises the cost of substitution a little, collects a little, and then watches its base fall as models get cheaper to run. It is a speed bump with a leak.
The real lesson is timing, not rates
Here is what the scenario gets right that the policy debate keeps missing. Both acts are debated in 2028 because nothing was measuring attributable output in 2026. The government in the story needs to transfer more at the moment it collects less, and it has no instrument ready. Our own Bill Gates piece made the same point about the Casar bill's unemployment trigger: a trigger is only as good as the meter it reads. The Casar bill reads the national unemployment rate. That is a lagging, economy-wide signal for a firm-level event.
A tax base is infrastructure. It takes years to build: definitions, reporting lines, audit practice, case law. None of that can be improvised in a crisis. Citrini's fictional administration reaches for a compute tax because compute is the only thing already metered. A legislature that wants a better option in 2028 has to build the meter in 2026, while the rate can still be zero.
Build the meter at zero percent
The campaign proposal, stated as policy design and not as law: require firms above a size threshold to report attributable autonomous revenue, the share of their revenue materially created or protected by autonomous systems, using the documentation standards they already use for transfer pricing and research credits. Set the contribution rate at zero. Publish the aggregate. Tie the rate to a trigger a legislature chooses later, firm-level payroll decline, sector displacement or a national unemployment threshold like the Casar bill's.
A zero-rate base costs the firms a reporting line and costs the public nothing. It passes the four tests. Base: revenue, already audited. Payer: the deploying firm, where the margin landed. Trigger: whatever the legislature writes, now that it has something to read. Erosion: none, because better models mean more attributable output, not less. And it passes the two 2026 gates, because it never asks whether the model was open or the app was cloned.
If the displacement in the scenario never comes, the rate stays at zero and the reporting line is a footnote. If it comes, the instrument exists on day one. Either outcome beats debating two bills in the middle of a bond market event.
Ask the next candidate which meter they would build
The scenario ends with the political divide deepening before an election. Ours is coming too. The useful question for any candidate is not whether AI should pay. It is which meter they would install now, at zero, so that the answer can be yes later without a crisis writing the bill.
One exercise for this week. Take the two acts and the Casar bill, and score each on the four lines: base, payer, trigger, erosion. Then score the zero-rate attribution base. Run one firm you know through the illustrative calculator below with the rate set to zero, and note the attributable revenue figure it produces. That figure is the meter. The rate is the politics.
