While we celebrate the nation’s workers this Labor Day, it’s a good time to remember that most Americans at any given time do not actually have a job. If you’re a child, or elderly, or a full-time caregiver, participation in the labor force may not be an option — or may just not even be the best use of your time. But, as columnist Matt Bruenig frequently reminds us, this reality creates unique challenges for societies that have chosen to allocate income through labor and capital ownership: Most people can’t put up much of either at any given time. And yet we want them to live thriving lives anyway.
This ever-present reality puts the risks of AI-induced unemployment into perspective. We already know how to take care of a society where most people aren’t working, because we are already forced to do it, however imperfectly: with unemployment benefits, free education, guaranteed health care, and other tried and true social welfare programs. Today, we are removing the paywall from Bruenig’s March essay in which he persuasively argues that, no matter how much unemployment you think AI will generate, “the solutions to economic inequality and insecurity have already been figured out, and productive technologies make them easier to implement, not harder.”
Whether or not we will do so is another matter entirely.
- John Thomason, senior editor
Prominent Democratic research firm, Blue Rose Research, recently released polling about the public reaction to artificial intelligence technology. Among other things, Blue Rose found that 54% of Americans think that unemployment caused by AI should be dealt with by “creating good-paying jobs” while only 17% favored providing “direct income support.” A further 49% favored a special tax on AI profiteers to finance these kinds of programs.
What’s actually interesting about this Blue Rose polling is not really the answers, but the questions and the underlying sentiment that we somehow need fresh new policy solutions to deal specifically with the negative distributive effects of AI.
In reality, the most boring, well-established social democratic policy approaches will work perfectly fine. That we have this kind of exotic discussion is emblematic of our deeply impoverished policymaking environment.
When thinking about the potential negative impacts of artificial intelligence, it is important to distinguish between the technology’s direct effects on things like production, education, and leisure (which may show up as AI cheating tools that hurt learning or AI companions that displace real social connection) and its indirect effects on things like the distribution of wealth, income, and consumption in society.
It is hard to know in advance what to do about the direct effects, both because we cannot really predict them and because each will likely require a specific, not general, solution. The opposite is true for managing the distributive impact of these technologies.
There is a well-established playbook for achieving economic security and equality that can be applied to an AI-infused economy just like any other advanced economy. Countries can levy taxes to fund universal welfare states in order to hold down income inequality and keep people’s material lives stable over time. They also can increase social ownership and impose targeted corporate taxes to put a lid on wealth inequality and excessive accumulation in specific industries.
The reason why the solutions to economic inequality and instability are always the same is because the problems are fundamentally the same. Capitalist economies distribute income by paying people for their labor and paying returns to people who own capital (wealth). Due to certain feedback loops, capital factors — or wealth — end up overwhelmingly concentrated in the hands of a small percentage of people. Labor factors — or work — are more dispersed, but individuals are constantly churning in and out of jobs, different jobs receive different pay, and a large swath of the population is not able to work at any given time.
Every AI-related distributive concern can be neatly organized into this basic framework.
Because it is a labor-saving technology, the implementation of AI throughout the economy will result in some people moving from employment to joblessness, at least for a time. But this is nothing new. It already happens in America over 60 million times a year.
Constant labor reallocation is a fixture of all dynamic economies. Yet it can be made relatively painless by establishing generous unemployment benefits and setting up a package of universal welfare programs — like child allowances, child care, education, and health care — that are not contingent upon one’s current employment or income.
Curiously, when it comes to differences in pay among laborers, the current conventional wisdom seems to suggest that AI would actually have an egalitarian effect. Right now, certain kinds of cognitive labor are compensated at much higher rates than other sorts of labor. This has resulted in a neat association of educational attainment with earnings, which has helped legitimize economic inequality as being fundamentally meritocratic in nature. But in a new technological environment in which the supply of “intelligence” that can execute certain cognitive tasks is massively increased, this association could attenuate or break down altogether, resulting in a wage compression that actually reduces earnings inequality among workers.
The most dystopian distributive concerns are focused on the capital side of the equation. There seems to be some fear that once AI technology is diffused throughout the economy, the handful of companies that control the frontier generative AI models — currently Google, Anthropic, and OpenAI — will hoover up massive shares of the national income each year. In this telling, most of the money companies save on labor will be parceled out to one of these companies, which will rapidly increase wealth and income inequality in the country.
Before getting into the solution to this hypothetical problem, it is worth dwelling a bit on how implausible this scenario actually is. For this to play out, a handful of companies will need to get themselves into a position where they can sell large language model (LLM) inference at an enormous markup. Yet, the LLM world is very competitive.
On OpenRouter, a service that makes it frictionless to switch between LLMs, there are currently 650 models — probably more by the time you read this. Many of these models are open source, meaning that, if you have a sufficiently powerful computer, you can run them yourself for free. Even if you lack such a computer, there is a highly competitive market of providers willing to sell metered access to those LLMs for a small fraction of what the frontier models charge.
Some of these open source models — like GLM-5, Kimi K2.5, and DeepSeek — are already at or near parity with the frontier models. And, given that it is possible to use frontier models to help train open source models, it seems likely that there will always be free models that are only a little bit behind the frontier.
What these dynamics point to is a future where LLM inference is largely commoditized, not a future of LLM super-profits. The tech sector is used to extracting rents through network effects and vendor lock-in. But not all tech products work like this, and, right now at least, LLMs are headed in that direction.
Of course, if this analysis is wrong and an AI-infused economy does result in a massive share of national income flowing to capital owners, especially for those selling LLM inference, this can be counteracted fairly trivially through social ownership of capital and corporate income taxes targeted at the LLM sector.
The most obvious vehicle for social ownership of this sort would be a dividend-paying social wealth fund like Alaska has. This is a fairly simple form of socialism but also pretty banal as far as things go. The federal government would just establish a new investment fund, fill it with stocks, bonds, and real estate, and then pay each American an equal share of the return on the fund every year.
This sort of proposal is often met with abstract objections about the mismanagement of capital and corruption. But the practical reality is that governments in the U.S. already run nearly 5,000 public pension funds that differ from this proposal only in that their return goes to fund retirement benefits rather than to fund a universal dividend.
There are many ways that such a fund could be financed, including wealth taxes, inheritance taxes, and even by selling bonds. In a world where the LLM sector began collecting massive profits, one obvious move would be to establish a special corporate tax just for that sector, with the revenues going to build the aforementioned social wealth fund.
This is something Norway has a long tradition of doing, especially for industries where it appears that excessive rates of return are being generated by companies that rely upon inputs that should, in the Norwegian way of thinking, belong to everyone. In the last few years, Norway imposed these special “resource rent taxes” on the nation’s aquaculture and wind energy sectors, the former because it uses the ocean as an input and the latter because it uses wind as an input. Prior to that, Norway took similar approaches to its oil and hydropower sectors.
This sort of taxation logic would apply equally well, if not better, to a mature LLM sector. In order to train their models, LLM companies use as inputs basically the entire corpus of human content creation. They aren’t just freeriding off unproduced inputs like petroleum or rivers, but on the produced inputs of the entire history of human meaning-making. If this type of enterprise generates super-profits, then we should simply do as Norway does and apply a special resource rent tax to the LLM sector in order to trim its net income down to size.
The fact that we have easy solutions to these problems does not mean that we will enact them.
But whether you buy those arguments or not, the AI policy debate raging right now presumes that the U.S. is on the verge of taking unprecedented action. Everything from nationalizing AI companies to a federal jobs guarantee seems to be on the table. If that’s the case, better to direct our energies to a big new program that will actually work.
A lot of early socialist writing shares a theme that making great scientific and technological advances is far harder than creating an egalitarian society. The idea was that, if human beings can figure out how to put a man on the moon, surely they can also figure out how to create a system in which nobody is poor and everybody can go to the doctor.
But the impediment to achieving egalitarianism is not technology or know-how. The solutions to economic inequality and insecurity have already been figured out, and productive technologies make them easier to implement, not harder. If, in the age of AI, the U.S. becomes even more of an inegalitarian nightmare than it currently is, it will not be because of AI, but rather because of the very same economic policy choices that plague us today.