AI and human work

The Work That’s Left

Ten years is not a long time. Go back that far in your head, find someone intelligent and curious about technology, and show them what you can do today. Pull up a blank text field. Type a sentence describing an image. It appears in three seconds, photorealistic, exactly what you described, without an artist, without editing software, without hours of work. Open a code editor. Describe the program you want in plain language and watch it write itself.

For that person in 2015, this technology was not interesting. It was the premise of a science fiction film that they are not sure they believe. You would watch the color leave their face.

For us, right now, it is a Tuesday afternoon. Routine. Nobody mentions it. The gap between what should feel like a miracle and what registers as background noise is not a detail about AI. It is the whole story.

The helper argument and where it runs out

The standard framing is: AI is a tool, the same as spreadsheets were, the same as the printing press was. History says that each wave of automation destroys some work and creates more elsewhere. Net positive. So the anxiety is misplaced.

This is not dishonest. It draws on real history. Every major automation wave did eventually generate more economic activity than it eliminated.

The word doing the most work in that sentence is “eventually.” The printing press took a century to reshape labor fully. The Industrial Revolution took generations. The workers displaced along the way did not live to see the net positive, and the current wave is moving through cognitive tasks at a speed no previous one has managed. The historical reassurance was always about the long run. The long run has a different meaning when the pace is this fast.

Also, every time a tool becomes transformative enough, it does not just change how fast existing work gets done. It changes what kinds of work exist at all. The typewriter did not make typesetters faster. It ended typesetting as a profession. The right question about AI is not whether it speeds up what you do, but whether the specific type of work you do remains necessary at all.

The canary is a programmer.

The clearest current case is software development, not because it is the most affected field overall, but because it is the most visible, and because there is an uncomfortable irony: the people who built AI are among the first to be displaced by it.

At the Council on Foreign Relations in 2025, Anthropic’s CEO, Dario Amodei, said the world might soon reach a point where AI is essentially writing all code. John Schulman, a co-founder of OpenAI, has publicly stated his median estimate for when his own role as an AI researcher will be fully automated: 2029. These are the people with the best view of what is coming.

The optimistic argument is that what survives automation is taste. The person who can evaluate a hundred AI-generated outputs and immediately know which one is correct, who understands the architecture well enough to catch a subtle flaw before it becomes a costly one, who has the judgment to recognize a good answer from a plausible-sounding wrong one. Execution gets automated. Judgment does not.

There is genuine truth in this. The programmer who can use AI tools well is more productive than ever before—the gap between them and someone who cannot has widened.

The problem is that this is a winner-take-all dynamic, not a stable middle ground. Tasks that once required a team of ten can now be done by two. The best people will not be replaced. There will be far fewer positions for anyone else. Good taste is a differentiator, but it is a differentiator in a room that is getting smaller.

The race with no finish line

In March 2026, economists Brett Hemenway Falk and Gerry Tsoukalas published a paper called “The AI Layoff Trap.” They modeled what happens when every company in a competitive market automates its workforce simultaneously, with full awareness of the collective result.

The conclusion: knowing the outcome is not enough to prevent it.

Their logic is this. When a firm automates, it captures the full cost savings. The workers who are laid off stop spending, reducing demand across the entire market. But that loss of demand is spread across all competitors, so each firm bears only a fraction of it. The firm that automates is better off than the one that does not, even though both firms are worse off than if neither had. It is a prisoner’s dilemma that plays out at a civilization scale, with rational actors driving toward a cliff that everyone can see.

In February 2026, Block cut nearly half of its ten-thousand-person workforce. CEO Jack Dorsey stated that AI had made those roles unnecessary and that most other companies would reach the same conclusion within the year. In 2025 alone, American employers announced over a million job cuts, with AI explicitly cited in roughly fifty-five thousand of them.

The paper looked at every commonly proposed response: universal basic income, retraining programs, worker equity, capital income taxes, and collective bargaining. None of them changes the firm-level calculus. Each automation decision remains a dominant strategy regardless of the surrounding support systems. The only thing that would change the math is a direct tax on automation itself, and no government has come close to that.

Why income support does not cut it

The UBI argument says: give everyone a floor, and the disruption becomes livable.

The floor is the problem.

Humans adapt to new conditions with extraordinary speed. It is one of the deepest features of how we work, the same mechanism that lets early humans survive in new climates, adjust to scarce food, and rebuild after disaster. We normalize the new condition and resume wanting more. Psychologists call this hedonic adaptation. Philip Brickman’s research in the 1970s showed that lottery winners and people who had been paralyzed in accidents both returned to roughly their previous level of reported well-being within a year. The size of the change did not determine how long the adjustment took.

Apply this to income. Whatever amount is set as a basic income floor will feel significant relative to the previous standard. Within a year, it is the baseline. Prices have adjusted because the businesses receiving that income compete for it, invest it, and bid up the prices of what they can buy. The new poverty threshold is the UBI number. Those at the very bottom are modestly better off in absolute terms and exactly as poorly off in relative terms, which is the terms that govern human psychology.

A billionaire who wants another billion is not a broken person. They have adapted to their current level, just as a person who grew up poor and now earns well has raised their threshold for “enough”. It moves with you. It always has.

Now consider AI. ChatGPT launched in late 2022. Within a year, generating a professional image, writing working code, getting a detailed legal summary, and doing all of it on a phone was something millions of people did without comment. What would have been incomprehensible in 2015 became routine in under three years. We have been living with mass-market generative AI for less time than the average lease agreement, and it already feels like it has always been there.

That is how fast we adapt. And whatever the world looks like after the automation Falk and Tsoukalas describe, we will normalize that too.

Moving faster than we understand

Buckminster Fuller observed in the mid-twentieth century that human knowledge had doubled roughly every hundred years up until 1900, then every twenty-five years by 1950. Ray Kurzweil formalized this pattern as the law of accelerating returns: the rate of technological progress roughly doubles every decade, which means that the amount of change in the next twenty-five years will equal the amount of change in the entire previous century.

A consequence of this, which nobody seems to agree on how to measure but everyone seems to feel, is that the gap between now and a point in the past far enough in the past that you would look like a god to those people keeps narrowing. It used to be centuries. The claim is that it may now be decades. By the time someone born today reaches old age, it may be only years.

Gloria Mark’s research at UC Irvine found that the average person’s focused attention on a single screen has dropped from 2.5 minutes in 2004 to 47 seconds in 2023. We consume roughly five times as much information daily as a person did in 1986, and retain almost none of it. One day after learning something, retention is around fifty percent. After a week, below fifteen.

The result is a kind of collective tunnel vision. Nobody is keeping up, individually or collectively. The rate at which the world is changing has exceeded the rate at which any person can form a coherent model of it. So we latch onto whatever is trending now, process it in fragments, and move to the next thing. No person or institution chooses the direction of travel. It emerges from billions of individual reactions to whatever is happening at this moment.

This is not unique to AI. It is the texture of the current decade. AI accelerates every dimension of it because AI accelerates everything, and that feedback loop has no natural ceiling.

What the question actually is

The wrong question to ask about all of this is whether AI will take your specific job.

It might. Some categories of work will contract, and others will not. The honest answer is that nobody’s models are reliable enough to tell you which is which, because the pace of change has outrun the economists, the technologists, and everyone else trying to predict it.

The more unsettling question is what is left when economic productivity is no longer the primary thing human beings offer.

For most of history, being a person meant being useful in some economy. Work was not just income. It was structure, identity, the social logic of how days were organized, and the answer to what you were doing with your time. If that changes, not for a few people at the margins but for a large fraction of the population within a generation, the disruption will not show up primarily as unemployment statistics. It will show up as what happens when the structure is gone, and no replacement has been agreed upon yet.

And we will adapt to whatever comes next. Faster than anyone expects. The world of 2040 will feel to the people living in it the way today feels to us, which is to say: normal, unremarkable, a new set of frictions to complain about. We are very good at this. It is maybe the only thing we are reliably good at.

Whether what we normalize will have been worth normalizing is the question nobody can answer now. It might be the only question that matters.