OpenAI Wants to Tax AI to Provide Everyone with a Subsidy. The Idea is Generous. The Conflict of Interest is Spectacular.
Sam Altman, who leads the company that is replacing cognitive workers with machines more than any other, proposes that the state tax AI and distribute the proceeds to everyone. The intention is good. Perhaps it could be more generous.
The Document and What It Really Says
On April 6, OpenAI published a thirteen-page document titled "Industrial Policy for the Intelligence Age." In an interview coinciding with the publication, Sam Altman described the scope of the ongoing transition as comparable to the New Deal and the American Progressive Era. The document proposes four concrete things: a robot tax, which means shifting the tax burden from labor to capital and the capital gains generated by AI; a National Public Wealth Fund that automatically allocates a share of AI industry assets to every American citizen; the experimentation of a four-day workweek at unchanged pay; and “portable” retirement accounts that follow workers from one job to another.
The proposal comes as OpenAI prepares for what is likely to be the largest IPO in recent history, with an estimated valuation of around $852 billion, and as it has just completed the transition from a non-profit entity to a for-profit corporation. It projects losses of $14 billion by 2026. The Trump administration has meanwhile signed an executive order limiting state regulation on AI. OpenAI opposes that regulation and simultaneously requests federal subsidies for its infrastructure. The document does not include specific rates for any of the proposed taxes. Critics have not overlooked these details.
The Structural Conflict of Interest
Here is a matter worth developing carefully. The Public Wealth Fund proposed by OpenAI would function like a sovereign fund similar to Norway's, but fueled by the growth of AI companies. The structural problem is that, in that model, the state becomes a shareholder of the same companies it should regulate. A government that finances welfare through dividends from the fund has a direct incentive not to touch the companies that generate the fund's growth. It becomes almost impossible to sanction an entity that directly contributes to citizens' income, or to tax it when necessary.
The financial logic leads to an outcome that commentators on the document have already noted: the proposal could crystallize the dominant position of current large AI players, turning the citizen from a worker with earned income into a passive recipient of an algorithmic dividend. Those who produce the dividend are no longer subject to regulation in the same sense they are today. On the political level, Greg Brockman (president of OpenAI) financed Trump's campaign. The current administration has reduced the corporate tax to 21% and is unlikely to return to different positions. The vagueness of the document regarding specific rates is not negligible in this context.
Layoff Data: What It Confirms and What It Doesn’t Say
The empirical picture of the labor market is more nuanced than the narrative from both sides suggests. Goldman Sachs published research last week estimating about 16,000 jobs eliminated per month in the U.S. directly due to AI, with the impact concentrated on entry-level roles, particularly in data entry, customer service, admin, and help desk.
Gen Z is the most affected in terms of volume, although Goldman itself notes that they are also the most technologically adaptive generation, with a higher probability of recovering positions with more analytical content compared to workers over 40 with specific skills and low mobility.
Using 40 years of data, Goldman tracked over 20,000 individuals using datasets from the Bureau of Labor Statistics. The most uncomfortable result for both narratives is that in the ten years following a technological layoff, real income growth is approximately 10 percentage points lower than for those who were never laid off, and 5 points lower than for those laid off for other reasons. The shock is not temporary. Moreover, researchers at Yale's Budget Lab are currently very cautious about the causal link: the 165,000 tech layoffs in the last year (from Microsoft, Amazon, Meta, Block, Oracle, and others) largely reflect a correction following the hiring boom during the pandemic and pressure on margins, not just automation. To complicate the picture further, a phrase that has become common again, AI washing, was publicly revived by Altman at the Indian summit in February: it describes the practice of publicly attributing layoffs to AI adoption when the real reason is something else, typically pressure on margins, because the narrative of technological replacement is more presentable to investors. The fact that this phenomenon is acknowledged by the CEO of the company selling AI suggests a level of caution in interpreting layoff numbers.
The Inflationary Mechanism: Where It Works and Where It Doesn’t
The most widespread argument against universal income is the inflationary one: if money is distributed to everyone, prices adjust and real purchasing power does not improve. This argument is correct for a specific category of goods, those with inelastic supply, especially real estate and land, where additional demand cannot be compensated by increased production. A homeowner knows that everyone has more liquidity, and the market adjusts quickly. On this part, the argument holds.
The question is that empirical pilots on universal income, in Finland, California, and Kenya, have not produced measurable inflationary spirals. They have instead led to reduced stress, maintained employment, and in some cases increased entrepreneurship. The problem is that these were small-scale experiments, in the range of thousands of participants, and scale matters a lot: a truly universal basic income has completely different aggregated mechanisms than a confined pilot. Evolving results without qualifications is a common methodological error in both directions of the debate. The more solid argument is not that UBI produces inflation across the board, but that for non-reproducible goods the absorption mechanism is almost automatic, and that on a national scale, the cost of even a modest universal income would require around 45% of GDP in the U.S. to guarantee a median salary, which makes the proposal harder to finance solely with AI profits.
The Real Problem: The Subsidy Shifts the Problem, It Doesn’t Solve It
If there is a structurally robust argument against the logic of universal subsidy as a primary response to automation, it is not the inflationary argument, but that of adaptation. Goldman data shows something specific: workers participating in vocational or technical training programs within three years of layoff earn about two percentage points of cumulative salary growth more in the following decade and have a 10 percentage points lower probability of returning to unemployment. Training is an empirical buffer with measurable track records, and the OpenAI document does not provide comparable data to support its distribution proposal.
This is not to say that the debate on redistribution is irrelevant: welfare relies on worker contributions, and if workers are structurally diminished, the system needs new sources of revenue. The automation contribution proposed in Italy by Stefano Bacchiocchi in the Senate on September 25, 2025, with an estimated revenue of 8 billion euros a year applied to large companies replacing staff with large-scale automation, is likely a more surgical approach compared to universal income, even if it carries competitiveness risks for medium-sized businesses and non-trivial implementation issues. The idea of taxing productive capital as an alternative to labor has a long history: Bill Gates introduced it into public debate with an interview with Quartz in February 2017, arguing that if a $50,000 worker is taxed, the robot doing the same job should be as well, and hypothesizing that the revenue could go towards retraining, schools, and elderly care. The European Parliament, a few days after that interview, rejected a similar proposal. In short, this is an ongoing debate, not an invention of 2026.
The point is that the work of adaptation, training, and shifting roles towards analytical and relational skills that current models do not replicate well has a practical advantage over subsidies: it requires much lower investments and has empirical evidence to support it. The Gen Z that manages to adapt does so because it has transferable skills on which it is easier to build something new, and the financial parachute alone, without the leverage of training, does not produce adaptation. Those trapped in a specific role without access to training are the real problem that neither OpenAI's robot tax nor the Public Wealth Fund seem to credibly address.
A Practical Test, on a Small Scale
There is a non-profit project that, at the end of March, almost at the same time OpenAI published its thirteen-page document, began distributing the first checks of a program that, in theory, does roughly what Altman proposes, only on a proportionally smaller scale. It is called AI Dividend and originated from two non-profit organizations, the AI Commons Project (a branch of the Fund For a Guaranteed Income) and What We Will, the latter founded by Kaitlin Cort, a software engineer and trainer who has seen the entry-level developer market emptying before her eyes as companies adopted Copilot and Claude, and the few remaining positions transformed into jobs for reviewing AI-generated code.
The program distributes $1,000 a month, with no spending restrictions, for a year, to an initial group of twenty-five to fifty people among developers, copywriters, journalists, call center workers, and data annotators, all with a clear connection between job loss and generative automation. The pilot budget is $300,000; the goal for 2026 is to reach $3 million, and achieving this requires new funders. Nick Salazar, who leads the Fund For a Guaranteed Income, states this clearly: negotiations with Anthropic are already open, and the point he raises publicly is that the CEOs of large labs claim to believe in supporting displaced workers, and if the checks do not arrive, it is legitimate to wonder how seriously they spoke. Sam Altman has actively supported the idea of UBI and has funded at least one structured experiment; Dario Amodei has called UBI "better than nothing"; Elon Musk has made statements in the same vein. At the time of writing, none of these three companies have opened their wallets for AI Dividend.
Cort and Salazar do not limit themselves to transferring cash: they pair payments with career orientation, and in some cases, the advice to beneficiaries is to move out of tech, shifting towards healthcare or skilled trades, as in those sectors the competition with current models is lower. This aligns with what the Goldman data suggested earlier as the most solid route, namely that training and reasoned mobility work better than mere cash transfers in the same direction of adaptation.
The Basic Income Earth Network, the academic organization monitoring basic income programs worldwide, has noted that AI Dividend is not a true universal basic income, as it is not universal but categorical, meaning it only goes to those who have lost their jobs for a specific reason. This distinction may seem pedantic, but it has practical implications, because a targeted subsidy cannot be used as empirical evidence for or against UBI, and American debate tends to confuse the two with a certain ease.
The Third Vertex of the Triangle
On the afternoon of April 21, while this article was being finalized, the Rockefeller Foundation announced a $100 million commitment over three years for a program called "Big Bet on Good Jobs for America." The stated goals are to create 1.6 million additional jobs nationwide, support 250 communities classified as distressed, and an estimated direct or indirect benefit for between 10 and 20 million people. The politically interesting part of this initiative is that this time it is not a government moving, nor a small non-profit waiting for a check from Anthropic, but a 113-year-old philanthropic foundation putting its own money into a problem that the AI industry has helped create, without waiting for players in that industry to decide whether they will contribute as well.
The detail that deserves attention is how the money is structured. It is not a subsidy, nor a redistribution of capital returns, but an investment in access to work, focused on four well-defined sectors: healthcare and care economy, energy transition, food supply chains, and AI-enabled industries. These are the sectors where current models do not replicate human work well and where aggregate demand remains structurally high. The parallel with the advice Kaitlin Cort gives to AI Dividend beneficiaries is evident and not coincidental: it is the operational version on a larger scale of the argument that the Goldman data already provided, namely that training and sector mobility yield measurable results in the decade following layoffs. It is also worth noting that three of the four sectors into which the Rockefeller is currently investing are almost exactly the jobs that Bill Gates, in 2017, indicated as natural recipients of revenue from a robot tax, with the difference that here the money is from the Foundation and arrives now, without waiting for someone to legislate a tax.
There is also a political detail worth mentioning. The head of the program is Derek Kilmer, who spent twelve years in Congress as a Democrat from Washington and is now senior vice president for the USA program at the Foundation. Rajiv Shah, the president, presents the initiative with a phrase that clearly distances Rockefeller from the OpenAI approach, stating that America does not work for everyone because too few Americans are working. This is not an appeal for universal basic income; it is a thesis on the centrality of the dignity of work even in the age of AI. The axis of reasoning is different: OpenAI’s proposal shifts the problem to the redistribution of capital returns, while the Rockefeller proposal shifts it to access to work in communities where jobs are scarce.
On the scale of projects, the comparison speaks for itself. OpenAI, poised to list at a valuation around $900 billion, has yet to put its own money on the table for the proposals in its document, limiting itself to announcing fellowships and research grants accompanied by API credits for those working on industrial policy ideas related to the theme. AI Dividend has a $300,000 pilot budget and is trying to reach $3 million with the help of Anthropic. Rockefeller is putting up $100 million as part of a total commitment of $300 million from 2023, and it is doing so now, not after the IPO.
The magnitude of difference between those who speak and those who pay is perhaps the most useful takeaway from these springtime news.
A Postscript on the Project's Name
The OpenAI document is called "Industrial Policy for the Intelligence Age" and cites the New Deal, the Progressive Era, and rural electrification of the 1930s as historical precedents. It does not mention that OpenAI is about to make the largest IPO in recent history. This informational asymmetry in the document is noted by various analysts, including AI Magazine.
The document exists. The conversation about the social contract in the age of AI is necessary. The proposals from those initiating that conversation, having around $900 billion of market capitalization at stake, merit careful reading of the missing details, not just those that are present. In the meantime, AI Dividend distributes $1,000 a month with a $300,000 pilot budget, while a 113-year-old philanthropic foundation puts $100 million on the table for a problem it did not help create, while the actors that have at least partly generated that problem continue to write documents instead of signing wire transfers.