OpenAI has just published a 13-page policy paper proposing robot taxes, a national public wealth fund, and a 32-hour working week. What makes the document unusual is the masthead. The company writing it happens to be the company most likely to cause the displacement the policies are meant to manage.

This is the same OpenAI that has accumulated roughly $600 billion in future spending commitments for data centres, projects $200 billion in annual revenue by 2030, and isn't expected to turn cash-flow positive before then. The company that has just minted seventy-five new multimillionaires through an internal share sale. The company whose products are named in the layoff announcements of half the Fortune 500.

When the people building a technology start writing the policy framework for managing its consequences, we should look closely at which consequences they're choosing to manage, and which ones happen to stay invisible.

The simplest way to read any policy proposal coming out of a dominant firm in its own industry is to ask whether the proposal would cost the company anything. A proposal with a real cost is responsibility; a proposal with no cost is positioning. Apply that test to the OpenAI paper and the picture sharpens fast — I'm not sure I've found a single one of the five proposals that fails it in OpenAI's favour.

What the paper actually proposes

The document is called Industrial Policy for the Intelligence Age: Ideas to Keep People First. Published on the 6th of April 2026, it sits somewhere between a white paper and a manifesto. Five proposals sit at its centre:

  • A public wealth fund modelled on Alaska's Permanent Fund, seeded by AI-company contributions, that would distribute returns to American citizens as a dividend.
  • A robot tax on companies using AI or automation to replace human workers, set at roughly the level of payroll tax the displaced worker would have generated.
  • A structural tax shift from labour to capital, justified by the prediction that AI-driven growth will hollow out the tax base funding Social Security, Medicaid and the rest if we keep taxing wages.
  • A 32-hour workweek pilot at full pay, framed as an "efficiency dividend" from AI productivity gains.
  • Automatic safety-net triggers designed to activate when AI-related displacement crosses a measurable threshold.

If a Scandinavian social-democratic party had published this paper, nobody would have blinked. These proposals are within shouting distance of mainstream European policy thinking. What makes the document unusual is its origin. OpenAI isn't a Scandinavian social-democratic party. It's a US-based capped-profit company with an implied valuation somewhere between $400 and $500 billion, and it holds the most concentrated power in the AI industry.

What dominant firms do when the heat rises

Heavily regulated industries have a familiar move. Once the disruption from a dominant firm gets big enough to attract serious political attention, the firm stops fighting regulators and starts writing the regulations itself. The drug industry did it with the Hatch-Waxman Act. The Master Settlement Agreement was tobacco's version of the same move; the banks shaped parts of Dodd-Frank along similar lines. In each case, the rules that emerged constrained those industries where the firms could afford to be constrained, and entrenched their position everywhere else.

Going through OpenAI's proposals one at a time, the answers are remarkably consistent.

Take the robot tax first. It falls on adopters, not on OpenAI. A bank that uses GPT-class models to automate a thousand customer-service jobs pays the tax. OpenAI sells the model and books the revenue. The cost flows through OpenAI's customers, not OpenAI's accounts.

Move to the wealth fund. It's seeded "in part" by contributions from AI companies, and those two words are doing a lot of work. A modest founder donation, a percentage of revenue at a level that doesn't bite — none of that would change OpenAI's trajectory. In exchange, the company gets to point to the fund as evidence of social conscience whenever the next round of displacement headlines arrives.

Now the labour-to-capital tax shift. It's a structural change to the US tax base. It would touch every shareholder in every public company, which is to say it would land on the people funding OpenAI's data centres just as much as anyone else. The shift wouldn't hit OpenAI any harder than it would hit Apple or Berkshire Hathaway, and it would arrive too slowly to disrupt OpenAI's funding picture in the near term.

The 32-hour workweek is a workforce policy aimed at everyone except the people working on frontier AI, who already routinely log 60-plus hour weeks because that's the culture of the firms doing the work. The pilot, by definition, isn't aimed at OpenAI's own staff.

And the automatic safety-net triggers sit downstream of the displacement itself. They activate after people have already lost their jobs. They don't constrain the rate at which displacement happens, and they pass the cost of managing it to the public purse.

None of these proposals would change OpenAI's strategic position. Several of them would reinforce it.

The financial subtext

A piece of this is about how OpenAI looks to the outside world right now, and that picture is more fragile than the public narrative admits.

The company has signed contracts worth somewhere close to $600 billion to buy data centre time and compute capacity. It generates around $2 billion a month in revenue. Internal projections seen by The Information suggest cumulative losses of $665 billion before the company turns cash-flow positive in 2030. Anthropic, its main rival, is targeting break-even as early as 2028.

The CFO, Sarah Friar, has told colleagues she doesn't believe OpenAI is ready for the 2026 IPO Sam Altman is pushing for. She wants 2027, partly because the books need cleaning up and partly because the maths get harder to defend if growth doesn't accelerate. Altman has reportedly responded by excluding her from several discussions on infrastructure and capital strategy. Bloomberg reported in April that OpenAI shares had grown almost impossible to unload on the secondary market as investors pivoted to Anthropic. The October 2025 employee tender at a roughly $400 billion valuation came in fine in the end, but the underlying volatility tells a story.

Sitting on commitments of that scale, with a profitability horizon of four years and a CFO openly worried about the IPO timeline, the cost of being painted as the firm that broke the labour market is real and immediate. Investor sentiment shifts. Political risk grows. New legislation, the unwelcome kind, written by people who weren't consulted, starts to look possible.

A 13-page policy paper that pre-empts the worst version of that legislation and positions OpenAI as the responsible adult in the room costs the company very little and buys it a lot. A firm with this much capital exposure and this little runway has every reason to shape the regulation it's about to face while it still controls the framing.

What's missing from the paper

The proposals are coherent. They address real problems. They borrow from genuine intellectual traditions. None of that's in question. What's worth noticing is what isn't in the paper.

Start with the deployment question. The paper has no proposal to slow the rate at which AI is being rolled out so labour markets and education systems can adjust. The pace of displacement is treated as a fact of nature, not as a choice OpenAI is making by shipping increasingly capable systems on an accelerating cadence.

Nothing in the paper costs OpenAI a meaningful share of revenue. Dario Amodei at Anthropic has floated a 3% token tax on language-model output, redirected to government for redistribution. The OpenAI document contains nothing equivalent. It would be remarkable if it did.

It also says nothing about how AI is sold or marketed to enterprise buyers as a cost-reduction lever. The narrative that frames AI as a way to "do more with less", which translates in practice to "do the same with fewer people", goes entirely unchallenged.

And it's silent on the disappearance of the entry-level work that creates future senior workers. I wrote about that disappearance in The graduation cliff: one in five companies have stopped hiring graduates because of AI, and the cost of that decision will land a decade from now, when the pipeline of trained mid-career people runs dry. The OpenAI paper has nothing to say about the pipeline question.

Then there's the reskilling assumption the paper leans on, the idea that displaced workers will be retrained into the new roles AI creates. That's the same narrative I argued against in The reskilling myth. The evidence for large-scale wholesale retraining is dismal. US federal programmes have been running since 1962 and have consistently failed to show statistically significant improvement in earnings or employment for participants. Upskilling within a domain works. Wholesale career pivots from assembly line to data scientist have never worked at scale, and there's no obvious reason AI changes that.

If the policy framework assumes a solution we already know doesn't work, the framework isn't doing the heavy lifting it claims to.

So is it responsibility or regulatory capture?

Both, probably. Sam Altman has talked about UBI and post-scarcity economics for the best part of a decade, and there's reason to think parts of the paper reflect actual convictions. There's also reason to think the timing — five months before the planned IPO, three weeks after Friar pushed back on the IPO date — isn't entirely coincidental. The sincerity question matters less than people think. What we should be asking is whether the regulation OpenAI is helping to shape would actually constrain what OpenAI does next. That's the test that should govern how the rest of us read this.

The proposals aren't bad faith. The 32-hour workweek would be a real social good if it actually happened; the wealth fund is a serious idea worth debating; the case for shifting tax from labour to capital has been made by economists across the political spectrum for years.

The problem is more subtle than bad faith. It's that when the company most affected by a regulation gets to set the agenda for what that regulation looks like, the regulation tends to look like what the company can live with. The proposals that would actually constrain OpenAI's behaviour — slower deployment, restrictions on how AI is sold as a labour-cost lever, mandatory revenue-share with displaced workers, a real seat for organised labour in shaping deployment timelines — are absent from the paper. They were never going to be in it.

A useful test for any policy proposal coming out of a dominant firm is the reverse test. Imagine the same proposal coming from someone with no commercial stake in it — a labour-side academic, say, or a union research department. Would the proposal look different? In OpenAI's case, almost certainly. A labour-side version of the same paper would include constraints on deployment pace, a meaningful self-imposed cost on the model providers, and structural reforms to enterprise AI procurement. The OpenAI version has none of these, and the omissions tell you whose paper it actually is.

What this means if you're running a business

If you're a CEO reading the OpenAI paper as a signal of where policy is heading, here's the read. The political conversation about AI displacement is moving from "should we worry?" to "what do we do?", and the people building the technology are working hard to make sure they get to frame the answer. The frame they're offering is one in which displacement is unavoidable, retraining mostly works, and the cost of the transition is borne by adopters, governments and the public, not by the model providers themselves.

You can adopt that frame or push back on it. If you adopt it, expect to be the one paying the robot tax when it eventually arrives. If you push back, the people you need to push back on are the AI companies whose policy teams are now writing the brief that policymakers will read, not the policymakers themselves, who are still catching up.

For most businesses the practical move is to stop treating AI deployment as a binary "use it or fall behind" decision. Roll it out in phases. Ask on a regular basis which roles are being made easier and which are being made to disappear. Track what's happening to the people whose jobs the technology touches. Ask whether the productivity gain is showing up in better products and services or only in a smaller payroll. These are the questions the next wave of regulation will eventually ask. You'll have better answers if you've been asking them all along.

The longer arc

A longer pattern sits underneath all of this. Every major technology that has reshaped the labour market, from agricultural mechanisation to industrial automation to digital platforms, has eventually produced a policy response. That response has usually arrived after the damage was done, written by people who weren't in the room when the technology was being built, and shaped by political coalitions the affected workers had to build from scratch.

OpenAI's paper is an attempt to write the response before the coalition forms. That isn't necessarily a bad thing — done in good faith and with real constraints on the company doing the writing, it could lead to better policy faster than the historical pattern would deliver. Done without those constraints, it leads to policy that looks progressive on the surface and entrenches incumbent position underneath.

The test isn't in the rhetoric of the paper. It's in what OpenAI is willing to give up. A robot tax the company actually pays, rather than passing through to its customers. A wealth fund seeded by a meaningful share of OpenAI's revenue rather than a token contribution. Constraints on how aggressively OpenAI deploys its own systems into the parts of the economy where displacement risk is highest. None of those are in the document, and absent them the paper reads as a positioning document dressed in social-democratic clothing. The history of regulatory capture suggests this is how it usually plays out, with the cost arriving slowly enough that nobody has to take responsibility for it at the time.

Quick answers

What is OpenAI's industrial policy paper? A 13-page document published on the 6th of April 2026, titled Industrial Policy for the Intelligence Age: Ideas to Keep People First, proposing a public wealth fund, robot taxes, a labour-to-capital tax shift, a 32-hour workweek pilot at full pay, and automatic safety-net triggers — all aimed at managing the labour-market consequences of AI.

Is OpenAI proposing these policies in good faith — and what's a robot tax anyway?

The sincerity question is interesting and largely a distraction. None of the proposals would meaningfully cost OpenAI revenue; almost all the cost would flow to OpenAI's customers, to other shareholders, or to the public purse. In OpenAI's version, the levy on automated work falls on adopters (the bank that uses GPT-class models to cut its headcount) rather than on the provider. Anthropic's Dario Amodei has floated a different version of the same idea — a 3% tax on language-model output, paid by the model provider, which would cost his own company a meaningful share of revenue. The two proposals look similar on the surface and very different in who pays. The article body explains why that matters more than whether Sam Altman believes his own paper.

What is regulatory capture?

When the firms most affected by a regulation end up writing it. The drug, tobacco and banking industries have all done it. The AI policy cycle is early enough that the pattern is still avoidable.

What should businesses do about AI displacement?

Stop treating deployment as a binary use-it-or-fall-behind decision. The next wave of regulation will assess AI deployments retrospectively — which roles were augmented, which were replaced, what the local employment effect was. Better to ask the questions now, while you still own the answers, than later, when whatever levy emerges arrives written by someone whose policy preferences you didn't help shape.