Six Ways to Pay for the Data AI Took — Scored

Washington is suddenly full of plans to share AI’s wealth. All of them are about future income. None settles the bill for twenty years of data already taken. Here is the full ledger — scored.

Returns to Data — Article IV of a Vector1 Research Program

By Brian Curry · Vector1 Research

The first three papers in this series established that data is a factor of production (I), that its share of AI value is roughly 45% while its payment is roughly 2% (II), and that the technology to meter individual contributions now exists (III). This paper turns from diagnosis to design — and begins by correcting an omission in the entire public debate, including our own earlier framing. The claim data’s owners hold is not one claim but two: a flow claim on future returns, and a stock claim on twenty years of data already taken and already converted into the most valuable companies on earth. We size both, score six mechanisms against a single measurable objective — how far each moves the Data Capture Ratio, Φ, at what cost — and show the mechanisms are not rivals but layers. The first layer is restitution.

1. The Debate Arrived. The Ledger Is Half-Written.

This month, the question this series has been building toward stopped being academic. On June 1, 2026, Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act, proposing a one-time 50% stock levy on the largest AI firms, flowing into a public fund with board representation. Months earlier, OpenAI’s own policy framework proposed a national wealth fund distributing AI profits to all Americans. The United Kingdom launched a £500 million Sovereign AI Fund in April. Everyone, suddenly, has a plan for sharing AI’s wealth.

Look closely at every one of these plans and you will find the same quiet assumption: the meter starts now. They are designs for dividing future income — next year’s profits, the next decade’s appreciation. Not one of them prices what already happened: roughly two decades of accumulated human writing, art, code, and knowledge, ingested at a price of zero, and already compounded into hundreds of billions of dollars of enterprise value. The only payment ever made against that history — Anthropic’s $1.5 billion settlement — was extracted by litigation, from one firm, for one corpus, and is treated in policy circles as an anomaly to be managed rather than what it is: the first installment on a debt.

A worker owed ten years of withheld wages does not call it justice when the employer offers to start paying properly from Tuesday. The public debate has been negotiating Tuesday. This paper writes down the whole ledger.

2. The Two Claims: Stock and Flow

Data’s owners hold two economically distinct claims, and conflating them has muddled every proposal on the table.

The flow claim is a factor payment: data’s ongoing share of the returns it generates, the γ̂ ≈ 45% of AI revenue measured in Article II, currently paid at four cents on the dollar. It is the claim our Data Capture Ratio Φ tracks, and it runs forward indefinitely.

The stock claim — the back-debt — is restitution: payment for the data already taken and already used. Every deployed frontier model is a compression of the 2005–2026 corpus, acquired for nothing. The back-debt can be sized two ways with the series’ own machinery, and the two ways bracket it. From below: apply γ̂ to cumulative AI-sector revenue to date, compounded at interest — back wages on the income the data already produced. From above: apply γ̂ to the enterprise value of the AI sector, since the unpaid input is capitalized in the valuations themselves — data’s share of the thing that was built:

The bracket is wide — on the order of $100 billion to $700 billion — and Article VI will tighten it. But even its floor reframes the entire policy conversation. The Sanders bill’s one-time 50% stock levy, usually argued as redistribution, is better understood as a creditor’s claim: a rough-justice collection of the stock debt. The bill’s weakness is not its size; it is that it never names the debt it is collecting, which leaves it arguing morality when it could be presenting an invoice. The settlement gave the invoice its first line item: ~$3,000 per work, times the corpus.

One claim looks backward, one forward; one is a lump, one a rate; one is restitution, one a wage. Any honest design must settle both — which is why the scorecard that follows has six rows, not five.

3. The Objective Function

A mechanism is worth adopting if the value it routes to data’s owners exceeds what it destroys along the way. Three terms capture this. The benefit is the value routed — for flow mechanisms, the increase in capture ΔΦ applied to data’s slice of AI revenue; for restitution, the share of the back-debt actually collected. Against it count two costs: administrative overhead — the bureaucracy, metering, and enforcement the mechanism requires — and incentive drag — the innovation and investment it deters:

This framing does real work, because the proposals on the table fail it in different ways. Some are cheap to run but capture almost nothing. Some capture enormously but at high incentive cost. One — restitution — has a property no flow mechanism shares: being one-time and backward-looking, it cannot distort future investment decisions at all, which is precisely why economists since Ricardo have preferred taxing windfalls already banked over taxing activity yet to happen.

4. The Six Mechanisms, Scored

The serious proposals reduce to six families. We score each on four criteria — capture potential, fairness to individuals, administrative feasibility, and innovation compatibility — on a 0–5 rubric detailed in the methodological note.

0. Restitution — the back-debt. Mechanism Zero, because it precedes the others logically and morally: settlement of the stock claim, through expanded class actions, a negotiated industry-wide settlement, or a statutory back-payment fund. Strengths: the largest single pool of value on the table; the legal machinery (class actions, claims administration) is a century old and battle-tested — the Anthropic settlement just ran it at record scale; and as a one-time payment on past activity it carries near-zero incentive drag. Weaknesses: tracing individual historical contributions is harder than metering future ones, so distribution will lean on rough proxies; and it requires either courage in Congress or a decade of dockets.

1. Licensing collectives. The status quo’s best version: publishers and platforms negotiate deals like Amazon–NYT ($20–25M/year). Strengths: market-driven, zero legislation, already happening. Weakness: it pays gatekeepers, not contributors, and only organized, lawyered corpora get paid. Ceiling: Φ ≈ 0.10.

2. Statutory royalty. The music-industry precedent: a compulsory license at a rate set by law, paid into a collection society — machinery that has run for a century at roughly 9–12% of revenue:

Strengths: proven plumbing, predictable, immune to holdouts. Weaknesses: politically set rates get set low; distribution favors registered rights-holders.

3. Sovereign data fund. The Sanders/Alaska/Norway family: a public fund paying universal dividends. Norway’s fund exceeds $2 trillion; Alaska has paid residents for fifty years. Strengths: universal coverage, high ceiling, compounds. Weaknesses: severs payment from contribution — the radiologist and the lurker receive the same check — and a 50% levy is at the aggressive end of incentive cost if framed as a tax on the future rather than collection of the past (§2’s reframe matters here).

4. Equity transfer. The Human Data Commons mechanism from our earlier work: data’s owners receive a stake — equity paying dividends and appreciation:

Strengths: the highest flow ceiling, because it captures appreciation, where AI value actually lives. It is also the natural currency for the back-debt: firms that cannot write a $200B check can issue a $200B stake, which is how every leveraged settlement in corporate history has worked. Weaknesses: hardest to legislate and administer; most exposed to constitutional challenge.

5. Attribution micropayments. Article III’s mechanism: meter individual contributions with in-run Shapley attribution and pay people what their data actually earned. Strengths: the only mechanism fair at the individual level, and the only one that creates supply of the scarce data the data wall demands. Weaknesses: the metering is eighteen months old; coverage starts narrow; and it is structurally forward-looking — it cannot reach the back-debt, which is exactly why it cannot stand alone.

Figure 1. Six mechanisms, four criteria, scored 0–5. No row dominates — and the weak cells don’t line up. Restitution scores strongly precisely where the flow mechanisms are weakest: it reaches the past, and it cannot distort the future.

Figure 2. Estimated Φ ceiling for the five flow mechanisms, against today’s 0.04 and the competitive benchmark of 1. Restitution does not appear on this chart by design: it settles the stock, not the flow — a one-time payment cannot lift a ratio of ongoing capture. That is not a weakness; it is a different claim.

5. The Result: It’s a Stack, Not a Contest

Here is the finding the scorecard forces, and it dissolves most of the current debate. The six mechanisms fail in complementary places — restitution reaches the past but not the future; licensing is feasible but shallow; funds are deep but unfair at the individual level; micropayments are fair but young and forward-only. Mechanisms with complementary failures are not alternatives. They are layers. And layers compound:

The flow layers compound toward full capture of the ongoing return; restitution is layer zero — the down payment that clears the ledger before the meter starts. The natural deployment order follows feasibility: licensing is already running; a statutory royalty can ride a century of collection-society plumbing; micropayments switch on as metering matures; a fund or equity layer captures appreciation — and the back-debt settlement, paid substantially in equity, can seed that fund, uniting the stock and flow claims in one instrument. The Alaska fund was seeded by exactly such a recognition: the oil under the ground already belonged to Alaskans before the first dividend was ever paid.

Figure 3. A phased deployment scenario for the flow layers, 2026–2036, plausibly carrying Φ from 0.04 to the 0.6–0.7 range within a decade. Layer zero — restitution — is not drawn: it is the lump-sum settlement of the stock claim that precedes and seeds the flows. The current policy debate — fund versus royalty versus market — is an argument about which layer to build first, mistaken for an argument about which to build at all.

This is the paper’s central claim, and we believe it is new: the data-compensation problem is two claims and a portfolio, and the proposals currently fighting each other in Washington are the portfolio — minus its missing first layer.

6. Why the Firms Will Come to the Table

A mechanism nobody adopts is a seminar. So the design must answer: why would AI firms ever agree to settle a nine-figure back-debt and pay 30, 40, 60 percent of data’s future return, when today they pay 4 and owe, officially, nothing?

Because the debt does not disappear by being unacknowledged — it accrues. The Anthropic settlement established that one corpus, at one firm, for past use, costs $1.5 billion, and every settlement resets the anchor for the next case upward. Discovery is getting easier, attribution tooling (Article III) now lets plaintiffs prove whose data is inside a model, and the nonrivalry multiplier means every new training run re-borrows the same corpus and extends the liability. The firm’s true choice is not “pay or don’t pay.” It is between a negotiated, capped, predictable settlement of the stock claim plus an orderly factor payment on the flow — and an open-ended, compounding, precedent-ratcheting legal exposure to both:

Figure 4. Cumulative expected cost to AI firms under the status quo (litigation exposure on the back-debt, anchored on the $1.5B-per-corpus precedent and scaling with cases) versus a negotiated path (stock settlement plus phased flow payments toward Φ ≈ 0.3). Litigation scales with corpora, plaintiffs, and years; settlement is capped and predictable. Past the crossover, paying is the cheaper strategy — and the rational firm moves before the crossover arrives.

To be precise about what this argument is and is not: litigation risk explains why settling is in the firms’ interest. It is not why the debt is owed — the debt is owed because the input was taken and the return was kept, the arithmetic of Articles I–II. A design that treated the back-debt merely as leverage to extract future flows would invite firms to buy amnesty cheap; the stock claim must be priced on its own terms (Equation 1), not traded away as a negotiating chip. The right structure is the one mass-tort and environmental law has converged on repeatedly: certainty for the defendant, fullness for the creditor — a statutory settlement that caps the firms’ historical exposure at a number reflecting the debt’s real size, paid substantially in equity, in exchange for releases and a clean meter going forward.

There is a second, quieter reason the firms come to the table, and it is the data wall from Article I. Firms are running out of free high-quality data anyway. A functioning payment system is not only a legal shield; it is a procurement channel for the scarce input. The first frontier lab to embrace the stack will not be conceding a tax. It will be cornering a market.

7. The Portfolio, Stated

Formally, the policy problem is to choose the mix of layers — restitution plus flows — that maximizes total routed value net of costs, subject to both sides staying in the game: firms preferring settlement to litigation, contributors preferring participation to opting out:

Our scoring implies a concrete sequence. Step zero (now): open the back-debt negotiation — a statutory settlement framework pricing the stock claim within the Equation 1 bracket, paid substantially in equity, with the Anthropic per-work figure as the floor of the per-unit price. Now, in parallel: extend licensing collectives and pass a statutory training royalty — lowest friction, proven plumbing, immediate Φ gain. Next (2–4 years): switch on attribution micropayments as in-run metering matures. Then (4–8 years): the equity from the back-debt settlement seeds the sovereign fund layer, which captures appreciation thereafter. In this light the Sanders bill and the OpenAI framework are both drafts of the fund layer — necessary, but each missing the ledger entry that justifies its own size: the bill proposes collecting roughly what Equation 1 says is owed, without ever saying owed.

8. Objections, Taken Seriously

“Retroactive payment is unfair — the rules allowed scraping at the time.” The rules never allowed it; they hadn’t priced it, which is different, and the courts are pricing it now — that is what $1.5 billion for past use means. Back pay for an unpriced taking is not retroactive rule-making; it is the ordinary work of restitution, the oldest branch of equity. And the economic version of the objection inverts: one-time payment for past activity is the least distortionary transfer available, because no future decision can avoid it.

“The back-debt number is unknowable, so it’s unworkable.” It is bracketed, not unknowable — Equation 1 bounds it from revenue and from value, and the settlement supplies a market-tested per-unit floor. Mass-tort law settles less tractable damages than this annually. Article VI tightens the bracket; the present point is that zero, the number currently penciled in, is the one figure everyone agrees is wrong.

“Stacking means stacking costs — six bureaucracies instead of one.” The layers share infrastructure: the registry restitution’s claims process builds is the registry licensing needs is the registry the royalty pool and the attribution meter need. The back-debt settlement, far from adding overhead, funds the plumbing the flow layers then reuse.

“A royalty or levy hands the AI race to rivals.” The objective function explicitly prices incentive drag, and the stock/flow split is what defuses this objection: the largest transfer (the back-debt) is one-time and backward-looking, with near-zero effect on future investment, while the flow layers are calibrated — and the data wall makes paying for data a competitive advantage, not a handicap. Resource markets did not end resource economies; they organized them.

“Universal dividends are fairer than attribution — everyone contributed.” Both intuitions are right, which is the deepest argument for the stack: the fund honors the diffuse contribution of everyone — and the back-debt, being largely untraceable to individuals, is exactly the claim a universal instrument should carry — while the attribution layer honors the measurable contribution of someone, going forward. Stock to the commons, flow to the contributor. The stack is the only design fair in both directions.

“This is politically impossible.” It was — until the largest copyright settlement in history, a Senate bill collecting an unnamed debt, a lab-authored wealth-fund proposal, and a sovereign AI fund all arrived within nine months of each other. The window where ‘impossible’ was a serious objection has closed. The question now is design quality, which is why this paper exists.

9. What Remains

Two papers complete the program. Article V — The Macroeconomics of a Data Dividend asks what actually happens when the stack is running: to growth, inflation, labor supply, and inequality, when the back-debt’s lump and the flow’s tens of billions start reaching households. And Article VI — The Transition maps the route: the post-settlement litigation landscape, the legislative vehicles, the tightening of the back-debt bracket, and the sequence of feasible steps from Φ = 0.04 — and a ledger showing zero — to the settled accounts this series has argued are owed.

The diagnosis is finished. The ledger is written — both sides of it. What remains is the consequences, and the route.

References and Sources

1. U.S. Senate (June 1, 2026). American AI Sovereign Wealth Fund Act, Congressional Record Vol. 172, №92. One-time 50% stock levy on largest AI firms; public fund; board representation.
2. Sanders, B. (May 2026). New York Times op-ed previewing the Act; “direct ownership stake” framing.
3. OpenAI (April 2026). Policy framework proposing a national AI wealth fund, automation taxes, and worker protections.
4. HM Government (April 2026). Sovereign AI Fund: £500M UK fund plus £282M strategic data-assets grants programme.
5. Fortune (June 2026). Sovereign-fund precedents: Norway’s >$2T oil fund; Alaska’s fifty years of resident dividends; 142,000+ AI-attributed layoffs in 2026.
6. Anthropic author class-action settlement (2025): $1.5B; ~$3,000/work — the per-unit floor for the back-debt and the litigation anchor in §6.
7. U.S. Copyright Office / Copyright Royalty Board: mechanical and statutory royalty machinery; historical rates ~9–12% of revenue.
8. Jones, C. I., & Tonetti, C. (2020). “Nonrivalry and the Economics of Data.” AER 110(9).
9. Wang et al. (ICLR 2025), “Data Shapley in One Training Run”; Zeng et al. (AAAI 2026), “Ripple Shapley” — the metering layer’s feasibility, and the attribution evidence available to plaintiffs in §6.
10. Vector1 Research (2026). Returns to Data I–III: the fourth-factor framework; γ̂ ≈ 0.45; Φ ≈ 0.04; the $34B/yr flow gap; the individual payment formula.

Methodological note: the 0–5 scorecard encodes Vector1’s structured judgment against stated criteria, not measured quantities; we publish the rubric so it can be attacked cell by cell. The back-debt bracket in Equation 1 applies Article II’s γ̂ to cumulative sector revenue (compounded, lower bound) and to sector enterprise value (upper bound); both inputs are public estimates and the bracket is deliberately wide pending Article VI. Φ ceilings are order-of-magnitude estimates from each mechanism’s structural coverage. Figures 3–4 are scenarios, not forecasts. Dollar figures nominal. Code and rubric: github.com/vector1-ai (forthcoming).

Brian Curry is a Kansas City–based data scientist, AI researcher, and founder of Vector1 Research. The Returns to Data program develops the economic theory of data ownership in the age of artificial intelligence.

Contact: brian at vector1.ai · License: CC BY 4.0

#ReturnsToData #DataDividend #Restitution #SovereignWealthFund #AIPolicy #MechanismDesign #DataOwnership #AIEconomics #BackPay #Vector1Research