The Bill That Was Never Going to Pass, and the Work It Does Anyway

Bernie Sanders published an op-ed in the New York Times this morning arguing that the American public should own half of OpenAI, Anthropic, and xAI. The bill is called the American AI Sovereign Wealth Fund Act. It will not become law. That is not the point, and understanding why it is not the point is the most useful thing you can do with this story.

The mechanism is equity transfer, not taxation. Sanders proposes moving 50% of equity in the largest US AI companies into a public fund, which would then pay dividends to citizens. Not a windfall tax. Not a compute levy. Not a licensing regime. Equity — the asset itself. This is a deliberate and consequential choice, and it creates three distinct effects that each deserve attention: a constitutional trap the labs cannot easily escape, a capital-markets disclosure problem that arrives whether the bill passes or not, and an Overton-window shift that makes narrower redistribution proposals look moderate by comparison.

The bill's most durable work happens in courtrooms and S-1 filings, not in committee votes.

Start with the citation strategy, because it is the sharpest move in the op-ed. Sanders notes that OpenAI previously floated a public wealth fund concept and that Anthropic publicly backed "national sovereign wealth funds with stakes in AI." He is, as he frames it, simply taking them at their word.

This puts the labs in an uncomfortable position. They can distance themselves from their own prior statements, which makes those statements look like the rhetorical positioning they probably were. They can stay silent, which reads as inconsistency. Or they can engage with the bill on its merits, which requires them to argue against a proposal they once appeared to endorse in principle.

None of these options is comfortable. All of them force a public record. Sanders' op-ed is an evidence-preservation exercise as much as a legislative one.

The "collective intelligence" framing is the philosophical engine underneath all of this. Sanders grounds the proposal in the argument that AI models are trained on knowledge and data produced collectively by the public, and therefore the public holds a cognisable claim on the asset. This is not a new idea. It echoes Jaron Lanier's data-as-labour arguments and the distributional work coming out of the RadicalxChange movement. What is new is a sitting US Senator mainstreaming it in a New York Times op-ed in the middle of an active IPO window.

The argument has real force as political rhetoric. Its force as legal or economic theory is weaker. Training on publicly available data is not the same as using publicly owned property in a legally cognisable sense; copyright law and the current case law on training data are contested but do not, at present, establish a public property interest in model outputs. "Collective intelligence" can carry the rhetorical weight Sanders places on it. It probably cannot carry the legal weight.

The constitutional question is real, and it runs in both directions. A forced transfer of 50% of privately held equity triggers Fifth Amendment Takings Clause (the constitutional provision requiring "just compensation" when the government takes private property) analysis immediately. This is not a technicality. No comparable mechanism has been used against any US industry. The government regulates; it can tax; it can, in extremis, condemn and compensate. Compelling the transfer of existing private equity without compensation is a different category of action.

The Washington Examiner's framing, "government take half of Anthropic and OpenAI," is adversarial but not inaccurate on the mechanism. The Takings Clause challenge would be filed within days of any passage, and the bill's sponsors almost certainly know this.

Here is the two-sided reading of that fact. The uncharitable version: Sanders is proposing something unconstitutional because unconstitutional proposals attract more attention and take longer to resolve, keeping the distributional question in public view. The charitable version: the equity mechanism is chosen precisely because it sidesteps the most obvious practical objection to redistributing AI wealth, which is that OpenAI, Anthropic, and xAI are not profitable. You cannot meaningfully tax profits that do not exist. You can hold equity in entities whose aggregate valuation is in the hundreds of billions.

Both readings can be true simultaneously. This is often how durable political proposals work.

There is a further practical problem the proposal underspecifies. A sovereign wealth fund holding 50% of illiquid private equity in companies that are not yet public pays no dividends until a liquidity event. OpenAI is mid-conversion to for-profit; Anthropic is privately held through ongoing funding rounds; xAI has not filed. Citizens would hold a theoretical stake in assets the companies themselves control the timing of realising. The distributional promise is fully contingent on IPO timelines and dividend decisions that remain in private hands.

Norway's Government Pension Fund (GPF), the most commonly cited model for sovereign wealth fund governance, took decades to build the institutional infrastructure to manage equity positions responsibly: voting rights protocols, conflict-of-interest frameworks, sector exclusion lists, dividend distribution mechanics. The American AI Sovereign Wealth Fund Act does not specify any of this. That is not disqualifying for a bill at this stage; first drafts rarely do. But it is a gap that any serious successor legislation would need to fill, and filling it is hard.

The IPO angle is the one that capital markets should be watching. Even a bill that dies in committee changes the disclosed risk environment for any AI company filing an S-1 (the registration statement a company files with the SEC before going public). Once this legislation enters the congressional record, any competent securities lawyer advising an AI lab on a public offering will flag it as a risk factor.

The risk factor does not need to say the bill will pass. It needs to say there is non-zero legislative appetite for equity-based redistribution mechanisms in the AI sector. That is now demonstrably true. Institutional investors conducting due diligence on an OpenAI or Anthropic IPO will price this probability, however small. At the valuations being discussed (OpenAI has been reported at valuations above $80 billion in private transactions; Anthropic's most recent funding rounds valued it in a similar range), even a small probability assigned to a forced equity transfer is material.

Sanders may do more work in the S-1 risk-factors section than in committee. This is not a metaphor. It is a plausible mechanism by which a dead bill affects live capital allocation.

The Overton window shift is the oldest move in the progressive legislative playbook, and it usually works. A 50% equity transfer is the maximalist position. Its existence creates political space. Democratic moderates who would never support 50% equity transfer can now support a compute tax (a levy on the processing power used to train and run AI models), an AI revenue levy, or an equity-grant requirement at IPO and position themselves as the responsible middle ground.

This is not speculation about Sanders' motives. It is a structural observation about how legislative anchoring works. The estate tax debates of the early 2000s, successive minimum wage proposals, and the Affordable Care Act's public option all followed this pattern: introduce the maximalist version, watch the maximalist version fail, use its failure to normalise something that would have seemed extreme before the anchor was set. The mechanism is well-documented.

Gizmodo's characterisation of Sanders as "the only Democrat(ish) lawmaker willing to govern on AI" is rhetorically useful but slightly misleading in this context. The more accurate description is that Sanders is willing to set the anchor. Other legislators governing on AI in narrower ways get more oxygen once the anchor exists.

What the labs called public benefit, Sanders is now calling a promissory note.

The strongest counterargument to the entire framing is worth taking seriously. It runs like this: Sanders' proposal, even as an anchor, may be too structurally alien to American property law to shift the conversation in the direction he intends. US legislative history on corporate redistribution is thin. There is no domestic precedent for equity transfer of this kind. European models (worker ownership requirements, mandatory equity-sharing schemes in certain jurisdictions) have domestic political and legal histories that do not translate. A proposal that looks maximalist-but-tractable in a European context may look simply alien in the US, which means the Overton window shifts less than Sanders hopes — or shifts in unpredictable directions.

This is a real constraint. It suggests the more effective anchoring strategy might have been a large compute tax or a mandatory IPO equity-grant requirement — mechanisms with at least some US precedent — rather than a sovereign wealth fund built on forced equity transfer. Sanders may have chosen the more dramatic mechanism because dramatic mechanisms attract coverage. That is a reasonable political choice. It is not the same as choosing the mechanism most likely to produce successor legislation.

The distributional capture argument underneath all of this is the one I find most durable. Whoever captured the value in the data the models trained on did not, in most cases, capture it voluntarily or knowingly. The concentration of AI-generated value in a small number of private entities is real and is accelerating. Sanders' proposal is an attempt to name that problem in legislative language and force a response. The specific mechanism may be constitutionally fraught, practically underspecified, and strategically mixed. The underlying problem it points at is none of those things.

The labs' own prior statements on public benefit, sovereign wealth, and broad distribution of AI gains are now in the record. Sanders put them there deliberately. Whatever happens to this bill, those statements are now fair game in every subsequent policy conversation about who AI is being built for.

That is not nothing. It might, in the end, be the whole point.

Footnotes

¶