The Sanders AI Fund Is Not Really About Money

Senator Bernie Sanders has proposed legislation that would force the largest US AI companies to transfer half their stock into a public fund, with a board, voting rights, and a roughly $1,000 annual dividend for every American. The headline number is $7 trillion. The headline mechanism is a sovereign wealth fund. Neither is the most interesting thing about the bill.

What the proposal actually does, if you read past the dividend, is name a question the AI debate has been carefully arranged not to ask: who owns the upside of a technology that was built, in large part, on public money and public data, and that is now being deployed in ways that move costs onto the people who built neither the models nor the companies?

The bill, briefly. The American AI Sovereign Wealth Fund Act would require AI firms above a revenue threshold to transfer 50% of equity into a federally controlled fund, governed by a bipartisan Independent Commission for Democratic AI, with members nominated by the president and confirmed by the Senate. The fund would hold voting shares and board seats. A 5% annual dividend would, on Sanders' office's numbers, work out to about $1,000 per American per year.¹

It is not going to pass. I want to say that early, because the legislative arithmetic matters less than what the proposal forces into the conversation. The US Senate has failed to pass comprehensive AI legislation of any kind. A bill that mandates the transfer of half the equity of OpenAI, Anthropic, and the AI divisions of Microsoft and Google starts well outside any plausible bipartisan window, and the Fifth Amendment's Takings Clause would invite immediate constitutional challenge to any compelled transfer without full compensation.²

So the question is not whether the bill becomes law. The question is what it makes visible.

What was already public. The foundation of every frontier model in production today rests on two public inputs that the current ownership structure treats as if they were free goods. The first is decades of publicly funded research — DARPA, the National Science Foundation, university labs whose grant lines made the underlying architectures possible. The second is the open internet: the text, images, code, and conversation of hundreds of millions of people, scraped without consent or compensation to produce the training corpora on which OpenAI's $300 billion valuation and Anthropic's $61.5 billion valuation now rest.³

This is not a fringe observation. It is the documented history of the field. What the Sanders bill does is take that history seriously enough to ask what follows from it. If the inputs were public, the standard property logic, we built it, we own it, does not cleanly apply to the outputs. The bill's claim is not that the public deserves a cut out of generosity. The claim is that the public already had a stake, and the current ownership structure obscured it.

You can disagree with the remedy and still find the framing hard to dismiss.

The dividend is the wrong number to argue about. Most of the coverage I have read this morning is fighting about the $1,000 figure. Is the $7 trillion valuation realistic. Would a 5% yield hold. Would the dividend be inflationary. These are real questions and they are also a way of not engaging with the structural argument underneath.

A dividend is a transfer. Transfers are politically legible — they look like welfare, or like a tax credit, and the existing debate knows how to argue about them. What a sovereign wealth fund with voting shares and board seats does is something different. It changes who is in the room when an AI company decides what to deploy, how fast, with what guardrails, and into which labour markets. That is a governance instrument dressed as a financial one.

This is where I think the bill is most serious and most under-discussed. The harms that are accumulating from current AI deployment — algorithmic wage-setting on gig platforms, warehouse surveillance systems that intensify the pace of work, hiring tools that filter applicants in ways the applicants cannot see or contest — are not problems that a cheque solves. They are problems of who decides. A fund with board seats is, at least in principle, a mechanism for converting public exposure into public voice.

A cheque compensates for harm. A board seat is a claim on whether the harm happens.

Where the bill is weakest. I want to be honest about this rather than wave it through. Equity ownership and governance authority are not the same thing, and the bill's design will determine whether the fund becomes a real check on deployment decisions or a passive index holder that collects dividends and rubber-stamps management. The Norwegian sovereign wealth fund holds stakes in roughly 9,000 companies and exercises governance influence on a narrow set of issues, mostly ESG-coded. That is a defensible model and it is also not what the displaced call centre worker in Manila or the paralegal in Ohio needs from public ownership of the firm that automated their job.

The Independent Commission for Democratic AI, as described, is nominated by the president and confirmed by the Senate. That is the same process that produces the Federal Communications Commission and the Federal Trade Commission, both of which have spent the last two decades demonstrating how thoroughly a nominally independent commission can be captured by the industry it oversees. If the commissioners are drawn from the same revolving door that staffs every other tech-adjacent regulatory body, the fund will own half of OpenAI and vote with management on every contested question.

The bill's principle is that the public should have a stake. Its design has not yet shown that the stake will be exercised on behalf of the public rather than on behalf of the firms the fund holds.

What the proposal is actually for. I think Sanders knows the bill will not pass in this Congress, and I think the bill is doing two things regardless. It is establishing a left anchor on AI ownership that did not previously exist in mainstream US politics, which shifts the centre of the debate even if the bill itself goes nowhere. And it is naming, in legislative language, a distributional question that the industry has been arguing should not be asked at all: whether the current allocation of gains and losses from AI deployment is a fact of nature or a policy choice.

It is a policy choice. The current arrangement, private equity, public exposure, no consultation with the workers most exposed, was not handed down. It was built, by specific firms, specific investors, and specific legislative inaction, over the last four years. Things that are built can be built differently.

The Sanders bill is not the answer. It might not even be a good answer. But it is the first US legislative proposal to take seriously that the question exists, and that is more than the rest of the Senate has managed since GPT-4 shipped.

Footnotes

¶