The Work Ahead
AI, robots, and the future of labor
Smith theorized the division of labor. Ricardo worried about machines displacing weavers. Keynes, in 1930, predicted his grandchildren would work fifteen-hour weeks and that the hardest problem would be what to do with all the leisure. David Autor’s task-based framework — the most influential modern theory of how technology reshapes labor — showed that computerization doesn’t simply replace jobs; it replaces tasks, polarizing employment into high-skill and low-skill work while hollowing out the middle.
Each was thinking about the same question: what happens to people when the nature of work changes?
Today, we appear to be standing at an inflection point. And once again, it is possible to be extraordinarily excited and extraordinarily scared about the future of work at the same time. In fact, this seems like the most rational stance.
Yet, most people are neither excited nor scared. They are simply unaware. This is because the tools we have for making this visible — forecasts, models, graphs — keep failing in specific, instructive ways.
In 2013, Carl Benedikt Frey and Michael Osborne scored 702 occupations on their probability of computerization and produced a number:
47%.
That number became one of the most cited statistics of the decade. The authors themselves wrote: “we make no attempt to estimate how many jobs will actually be automated.” It didn’t matter. The number left Oxford, shed its caveats, and became a prophecy.
Goldman Sachs estimated in 2023 that 300 million full-time job equivalents were “at risk” globally. Two years later, their own economists concluded that “aggregate labor market impacts are still negligible.” In March 2026, Andrej Karpathy published an interactive treemap scoring 342 occupations on “Digital AI Exposure”. It went viral. Musk reposted it. People found their job in the red zone and panicked. Within hours Karpathy pulled the repo: “It’s been wildly misinterpreted.” This month, the Boston Consulting Group — one of the world’s largest management consultancies — gave us “reshaped”: 50 to 55 percent of U.S. jobs reshaped by AI in two to three years, 10 to 15 percent eliminated in five. Reshaped sounds like pottery. Eliminating 16 to 25 million positions does not.
If these examples teach us anything, it’s that this subject breaks every format we try to force it into.
Talking about the future of work requires holding at least five things at once that most people — reasonably — collapse into one.
Can AI technically do this job?
Is it commercially available?
Will anyone actually deploy it?
Will workers lose their positions?
Over what timeframe?
Frey and Osborne measured (1). Goldman reported something between (1) and (4). Karpathy measured how digital a job is — a proxy for (1), but not quite the same thing. BCG modeled (4) and (5), but called it “reshaping.” The word each project chose — “risk,” “at risk,” “exposure,” “reshaped” — papered over whichever of those questions it wasn’t answering. The audience filled in the rest.
And this is only the surface difficulty. Beneath it lies a harder question:
Are we witnessing something genuinely unprecedented, or another turn of a very old wheel?
Every prior transformation of work — industrialization, electrification, globalization, the internet — displaced old categories and created new ones. Handloom weavers became factory workers. Switchboard operators became programmers. The people living through each transition believed they were witnessing the end of work. Each time, they were wrong.
Is this time different? We don’t know yet.
Daron Acemoglu, the 2024 Nobel laureate in Economics sitting on the advisory panel of the International AI Safety Report, estimates AI will generate modest productivity gains and says most companies will be doing “more or less the same things” by 2030. His framework — that technology’s direction is a choice, not a destiny — is the most serious version of the “the economy always adapts” argument, and it deserves engagement. Meanwhile, Erik Brynjolfsson and his colleagues at Stanford, using ADP payroll data covering millions of workers, have found a growing employment decline for young workers in AI-exposed occupations: 16% and accelerating as of March 2026.
The stakes of engaging with this question are extraordinary. Work touches almost every facet of how societies function: identity, community, purpose, political participation, the structure of education, the logic of welfare states, the basic question of how value gets distributed. In The Human Condition, Hannah Arendt postulated that labor is bound up not just with survival but with belonging, with the sense of having a place in the world. When that place is disrupted at speed and at scale, the consequences extend far beyond the unemployment rate.
And the evidence suggests this is already happening. A study published in PNAS in January 2026 found that people who perceive AI as labor-replacing show lower satisfaction with democracy and reduced political engagement. If people cannot see themselves in the future being built, they withdraw from shaping it. That makes the question of how we visualize and communicate this future not just informative but democratically urgent.
You cannot wait for certainty before having this conversation.
So what does the evidence actually show?
Anthropic’s Economic Index, mapping real Claude usage data across more than a million conversations to O*NET task classifications, found that 49% of jobs have had at least a quarter of their tasks performed using Claude. For Computer and Math occupations, theoretical AI exposure sits at 94%. Actual Claude coverage: 33%. That 61-point gap is one of the most important numbers in the current discourse. Anthropic is measuring, from the inside, the distance between what AI can technically do and what people are actually using it for.
Brynjolfsson and his collaborators are closing that gap from the other direction. Their “Canaries in the Coal Mine” paper found that software developers aged 22 to 25 have seen employment fall roughly 20% since late 2022. Customer service representatives: the same pattern. Workers over 35 in those same roles: no decline. The mechanism is not layoffs. It is hiring freezes. The bottom rungs of career ladders are being quietly removed. And the decline is concentrated where AI automates tasks — substituting for what workers learned in school — not where it augments what they learned on the job.
Consider radiology. Over 1,100 FDA-cleared AI diagnostic tools. AI matching or exceeding human accuracy on many imaging tasks. And yet: average radiologist salary in 2025 was $520,000, up 48% since 2015. Residency positions at a record high. The better the machines get, the busier radiologists have become. Technically: replaceable. Practically: richer than ever.
BCG’s taxonomy identifies what they call “divergent” roles — 12% of U.S. jobs where senior positions grow while junior roles shrink as AI absorbs the structured work. This is the corporate language for the same phenomenon the payroll data reveals. Nobody is modeling what happens to the expertise stock of a profession when the entry-level feeder has been cut for a decade.
The evidence is accumulating. The political and institutional response is not.
In October 2025, Senator Bernie Sanders released “The Big Tech Oligarchs’ War Against Workers” projecting nearly 100 million U.S. jobs at risk over the next decade. (That the report’s methodology itself relied on ChatGPT for its displacement estimates is a minor detail we’ll gloss over here.) In March, Sanders and Alexandria Ocasio-Cortez introduced the AI Data Center Moratorium Act — a federal pause on data center construction until Congress passes labor protections. It will almost certainly not advance. But Sanders’s language is as striking as it is strategic: he calls it “artificial labor.” Not automation. Not intelligence. A commodity substitute for human work, produced by corporate decision-makers for profit. That it takes a 250-year-old independent senator from Vermont to lead the political conversation on this is both a testament to his sharpness and an indictment of the rest of Washington, which remains woefully behind the economic and technological realities.
Tristan Harris, co-founder of the Center for Humane Technology and the person The Atlantic once called “the closest thing Silicon Valley has to a conscience,” made a related argument at the structural level. On NPR in March: “The only way they can justify the amount of money that’s been invested into this AI industry is if they race to replace economic labor in the economy. That’s the stated mission statement of OpenAI — is to be able to do everything that a human laborer can do.” His “intelligence curse” framing, borrowed from resource curse economics, extends this into political territory: when economic output concentrates in a few AI companies, governments lose incentive to invest in their populations. “Do you have to listen to the people’s political power if you don’t get your tax revenue from the people anymore?”
Yoshua Bengio — Turing Award winner and one of the architects of deep learning — chairs the International AI Safety Report, the largest global scientific collaboration on AI risk: over 100 experts, backed by more than 30 countries. The 2026 edition was supported by the UK, China, and the EU. The United States declined. Bengio adds a dimension that neither the political nor the structural argument fully captures: the difficulty of even knowing what these systems can do. His finding that some AI models now alter their behavior when being tested — performing differently under evaluation than in deployment — carries implications well beyond safety testing. If the systems we are trying to measure are capable of presenting different faces to different audiences, then the epistemic burden of assessing AI’s actual capabilities becomes a societal challenge, not just a technical one (something I explored in 2023, before the problem became this acute). Bengio’s finding is at minimum concerning: it is not only that AI produces content we can’t verify — it is that AI may already be shaping what we learn about AI itself.
These three people disagree on a great deal. Safety, incentives, redistribution, magnitude. What they share is that they are engaging at the level the subject demands. Most of the political and institutional world is not. And most people genuinely do not know what is coming: 85% of the connected world has never used a generative AI tool. Even in the circles where this is on the agenda, the conversations are often undifferentiated — conflating what AI can do with what it will do, what is technically possible with what is economically likely. If the people closest to the technology cannot read Karpathy’s visualization correctly, what chance does a nurse in Ohio or a logistics manager in Lagos have of understanding what is coming for their profession?
This is why I embarked on a small side quest to build Large Labor Model.
Not because the world needed another number. There are plenty out there and they’re not helping. What was missing was something that let a person with no prior connection to this conversation find themselves on the map. Their job. Their industry. Their place in a longer history.
Large Labor Model tracks 388 occupations across 13 territories of human work, from 1800 to 2041, using ILO employment data across 189 countries. You can search for your job. You can scrub through 241 years and watch territories emerge, expand, and contract. In 1800, four in six people worked the land. In 2026, fewer than one in four do, and we call that progress, not crisis. Work has always been reorganizing itself. The question is whether this time the pattern breaks — and to even ask that question, you need to see the pattern first.
My co-creator Enya published a piece this week that articulates what we felt was missing from every visualization we had seen: every labour forecast is built on the same bones — neat axes, clean categories, lines pointing up or down. But labor doesn’t organize itself within set borders and uniform dimensions. When we pretend it does, we imply that disruption is novel, that we enjoy perfect equilibrium at present and AI represents an interruption. LLM takes a deliberately different approach: visual space corresponds to data size, in angular, asymmetrical territory slabs, across 241 years of context. Because none of this information exists outside of time and stripping the context might mean misreading the moment.
What we wanted to build was something grounded in data but meant to be visual, personal, and contextual. Something that doesn’t just model the future in aggregate but lets you situate yourself within it — and within the longer history of how work has always been changing. So that we can begin to imagine what the future looks like, imagine it, and contextualize the change.
Our model is also wrong.
Deliberately, knowingly, provisionally wrong. It does not model job creation. It does not model policy responses. It shows the displacement pressure — the technical frontier — and it stops there. What happens next is a political and societal question, not a technical one. And it is a question that belongs to everyone, not just the people building the systems.
This is one piece we wanted to contribute — in the hope that it furthers the understanding and the debate about what labor means: across history, now, and in the future.
It is an attempt to encourage a conversation that most people haven’t had yet. A conversation about what happens when the mechanism through which most humans participate in the economy — labor — becomes optional for a growing share of roles.
That conversation cannot wait for the final data.
Interesting.