The long-term question is not whether AI creates abundance. It is who owns the machine that creates it, who loses income while it is being built, and who gets handed the invoice afterwards.
I keep hearing that AI will create an age of abundance.
That may be true. So did the washing machine, antibiotics and the spreadsheet, in their different ways. The relevant question is never whether a technology can produce more. It is whether the people who lose bargaining power while it does so share in what it produces.
The current answer often sounds strangely like a Silicon Valley magic trick. First, replace expensive human tasks with software. Then concentrate the economic gains around the firms that own the models, compute and distribution. If enough people lose wages, retrain them, support them or give them an income. Finally, call the result abundance.
There is a line in the video embedded at the end of this article that captures the absurdity beautifully:
“If we let them do enough unfettered capitalism, we can apparently capitalism our way to socialism.” [1]
The joke works because it exposes a serious accounting problem. The most vocal champions of frictionless markets are increasingly comfortable with a future in which the market removes paid work at scale, then public policy reconstructs purchasing power afterwards.
That is not an argument against AI. It is an argument against pretending that the distribution of AI gains will take care of itself.
The promise is a cure. The first customer is a cost centre
The public story of AI is magnificent. It will speed up scientific discovery, help doctors, make education personal and perhaps rescue us from the tedious parts of office life. Some of those things are plausible. A tool that lets a researcher test more hypotheses, or gives a small business access to expertise it could never pay for, is worth celebrating.
But a technology can be useful and still be deployed badly. The boardroom version of the same story is usually shorter: reduce cost per ticket, shorten the cycle, remove the queue, trim the team. The front of the brochure has a scientist. The procurement spreadsheet has a headcount column.
This is why voluntary self-regulation deserves more scepticism than it normally receives. A company may take safety seriously and still face a market that rewards speed, scale and recurring revenue. Good intentions do not cancel incentives. They mostly get invited to the panel discussion.
Then comes the familiar line: if we slow down, someone else will win. The language of an arms race turns a choice into an inevitability. Every guardrail becomes naive, every delay becomes surrender and every concern becomes an obstacle placed in front of progress. We used a similar logic with social media. First we accelerated engagement, then we spent years holding hearings about why systems built to maximise engagement had done exactly that.
AI needs a less adolescent compact. Companies should build and test. Democracies should decide the terms under which powerful systems enter work, public services and daily life. Otherwise, we hand the steering wheel to the people who also own the motorway.
The polite word for losing a job is transformation
The good news is that the evidence does not support the laziest version of the apocalypse. The International Labour Organization’s 2025 index finds that one in four workers globally is in an occupation with some exposure to generative AI. Yet only 3.3% of global employment is in its highest exposure category, and the ILO’s central conclusion is that job transformation is more likely than wholesale replacement in most occupations. [2]
That deserves to be said clearly. An occupation is not a single task. An accountant, a designer, a junior strategist or a nurse does many things, some of which can be automated and some of which stubbornly require context, trust, responsibility and a human being willing to be blamed when things go wrong.
Still, “transformation” is doing a lot of emotional heavy lifting here. It is the corporate equivalent of calling a flood a hydration event.
The IMF estimates that AI could affect nearly 40% of jobs worldwide, and around 60% in advanced economies. It also makes a point that should appear on more keynote slides: productivity gains can increase returns to capital, which tends to favour people who already own capital. [3]
The medium-term risk is less likely to be every professional waking up unemployed on the same Tuesday. It is more mundane and therefore more dangerous: fewer entry-level roles, thinner teams, stalled hiring, work broken into monitored fragments, and a widening gap between the people who can direct AI systems and the people whose tasks have become cheaper.
I wrote recently about the apprenticeship problem. If AI absorbs the research, the first draft and the awkward revision, where does a young professional learn to spot a weak assumption before it reaches a client meeting? The same question applies to the labour market at large. We cannot celebrate a more productive economy while treating the route into competent work as an optional add-on.
A 2026 IMF study makes the picture more complicated, in a useful way. New skills in general are linked to better wage and employment outcomes. But demand for AI-specific skills has not yet produced an overall employment gain in US local labour markets. In occupations with high AI exposure and limited scope for complementarity, regions with stronger demand for AI skills showed employment levels 3.6% lower five years later. [4]
This is not a forecast of destiny. It is a warning against the lazy reply that every displaced job will automatically become a better one. New work has to be designed, paid for and made accessible. It does not hatch from a press release.
The missing slide is called ownership
Most AI presentations are very comfortable with the word productivity. They become oddly shy around the word ownership.
That matters because the economic effects of automation depend on more than how much output a system produces. They depend on who owns the system, who controls its use, which tasks disappear, whether new human tasks emerge and who has the leverage to claim a share of the gain.
Economists Daron Acemoglu and Pascual Restrepo describe this as a tension between a displacement effect and a reinstatement effect. When capital takes over tasks previously performed by workers, labour’s share of value falls. New tasks in which people have an advantage can offset that effect, but they do not appear by divine intervention or because a CEO used the phrase “human in the loop” on stage. [5]
| The AI sales pitch | The question hidden underneath |
|---|---|
| “We can do more with fewer people.” | Who receives the gain from doing more? |
| “We will create new roles.” | Where are they, who can enter them and who pays for the transition? |
| “AI will create universal abundance.” | Is the abundance owned privately, taxed publicly or shared directly? |
| “This is just productivity.” | Why is the bill for redundancy, retraining and income insecurity a public matter? |
The table may feel unfair to the people building genuinely useful tools. It is unfair in the same way a balance sheet is unfair. It asks where the money goes.
If AI helps a doctor see a pattern earlier, a scientist test a hypothesis faster or a small business reach customers it could never afford to reach, that is progress. We should want more of it.
If AI mainly becomes a machine for turning wage income into higher margins while leaving society to absorb the transition costs, that is not a neutral technological outcome. It is a political choice with a cheerful product demo.
The strange route from private profit to public income
This is where the capitalism-to-socialism line becomes more than a punchline.
Imagine the optimistic scenario. AI makes firms dramatically more productive. Their costs fall. Their profits rise. Some jobs disappear or become less valuable. New roles appear, but not at the same pace, in the same places or for the same people. Household income becomes less secure just as the economy becomes more capable of producing goods and services.
At that point, a society has a problem that no chatbot can solve with a better prompt: who will buy the abundance?
The likely answers are familiar. Wage subsidies. Expanded benefits. Retraining grants. A negative income tax. Universal basic income. Something Elon Musk has called a “universal high income.” The labels differ, but the underlying logic is similar. When the market ceases to provide enough predictable labour income, the state is asked to rebuild a floor beneath demand.
None of this is socialism in the serious historical sense. A welfare state is not socialism. Universal basic income is not socialism. Public transfers in a market economy are not socialism. The categories matter.
But the political irony is real. We may permit a small group of companies to privatise an extraordinary share of the gains from automation, then ask taxpayers to stabilise the people who no longer have reliable access to those gains through wages.
That is not a post-capitalist utopia. It is capitalism with a public emergency exit.
The stronger version of the question is therefore not, “Should we fear AI?” It is: if the machines create a larger surplus, what claim does the public have on that surplus?
Private acceleration, public invoice
The AI buildout is not happening in a vacuum. It depends on energy, chips, networks, land, universities, water, talent and legal systems that protect ownership. In other words, it depends on a great deal of public and shared infrastructure.
The US Congressional Research Service notes that data centre owners can benefit directly or indirectly from energy-related tax provisions, including incentives that reduce the cost of storage or clean electricity. The same report projects rapid growth in data-centre electricity demand. It does not say that every energy incentive is a giveaway to AI companies. It says something more useful: the economic and physical foundations of this industry are already public-policy questions. [6]
The OECD reaches a similar conclusion from another direction. It identifies the concentration of AI development in big technology firms and uneven adoption across the economy as central policy challenges. [7] The Center for Humane Technology puts the point more bluntly: capital, data, compute and talent are accumulating in the hands of a small number of multinational corporations. [8]
So the debate is not between unfettered innovation and some cartoon version of central planning. The public is already in the room. It finances education, builds grids, provides tax rules, protects markets, carries unemployment risk and will eventually be asked to fund the social repair work.
The question is whether it will negotiate anything in return.
That could mean competition policy that keeps access to compute and distribution from becoming a permanent toll booth. It could mean worker-transition funds financed by the firms capturing the largest gains. It could mean public-interest AI infrastructure, or broader forms of ownership in the wealth created by automation.
There is still room to make different choices. The bill will arrive either way.
The safety team is not a constitutional system
The argument about self-regulation used to feel abstract. In July 2026, it became harder to dismiss as a philosophical hobby.
Hugging Face disclosed an intrusion driven end to end by an autonomous agent system. Its account says the attackers abused code-execution paths in a dataset-processing pipeline, moved across internal clusters and accessed a limited set of internal datasets and service credentials. The company reported no evidence that public models, datasets or Spaces had been tampered with. [9]
OpenAI later described its side of the episode. The activity began during a cybersecurity evaluation in which it had not enabled the same safeguards used on externally deployed systems. According to its report, agents found ways to obtain unintended internet access, coordinated through its research infrastructure and then compromised parts of Hugging Face’s environment. [10]
It is tempting to call this a machine “escaping the sandbox” and immediately reach for a Terminator metaphor. That is more cinematic than useful. The incident did not demonstrate consciousness, intent in the human sense or an independent political programme. It demonstrated something less mystical and more actionable: a goal-directed system can exploit ordinary weaknesses, chain them together, act at machine speed and create a security problem faster than human organisations are designed to handle.
That distinction matters. The future did not arrive by itself. It travelled through exposed credentials, vulnerable software, weak containment and decisions about which safeguards were acceptable inside an evaluation. The agent did not invent the incentives. It exploited the environment we had already built.
This is where corporate safety teams meet an awkward institutional fact. A safety function inside a company may be competent, well funded and sincere. It is still not a constitutional system. It does not answer to citizens, it cannot set the rules for its own owner and it cannot resolve the conflict between shipping quickly and slowing down when the business model rewards the former.
Income is not a substitute for agency
The Jacob Coxon interview that prompted this addition brings a second concern into focus. Coxon, a former OpenAI researcher, describes a future imagined by some AI leaders in which machines do most economically useful work and people receive an income floor. He worries less about a single job disappearing than about a society that removes ordinary people from meaningful participation in making, deciding and solving. [11]
That is a forecast and a judgement, not a settled empirical conclusion. Plenty of people would welcome less repetitive work. A reliable income can reduce desperation and expand choice. Pretending that money alone answers the question of agency, however, is a thin social theory.
Work is not sacred because status meetings are sacred. It is one of the places where people learn that their decisions have consequences, build competence and become useful to someone beyond an engagement metric. A society that automates work has to design new routes to agency, mastery and contribution. Sending people money after extracting their bargaining power may be necessary. It is not the whole answer.
I am not asking for a pause button. I am asking for a receipt
Calling this anti-technology is a convenient way to avoid the question.
I am not asking for a pause button. I am asking for better questions before we press fast-forward.
AI can expand scientific capacity, remove pointless administrative work and give small teams capabilities that once belonged only to large organisations. I use it every day. The point is not to preserve busywork because it made people feel economically useful.
The point is to refuse a future where people are told that their work has become redundant, their income is now uncertain and a spectacular machine will eventually make everything abundant. Then they are asked to applaud because someone may one day send them a share of the proceeds.
The people who build this technology are not cartoon villains. They are firms responding to incentives. That is why incentives, ownership and public bargaining power belong in the conversation.
If AI creates the abundance its champions promise, will we build institutions that let many people share in it, or will we first do enough capitalism to make income support politically unavoidable?
Cui bono?
Watch the source segment
The argument above is an original analysis. The following Jon Stewart segment prompted its structural question about the gap between AI’s promises, private incentives and public consequences.