Notes on labor, value, and purpose.

Dario Amodei (Anthropic) has sketched a future in which economic growth reaches 10-15% annually while unemployment rises to levels the modern West has never sustained. The two claims sit uncomfortably in the same sentence. For most of the modern era we have assumed that growth and employment moved together. If output was expanding it was because more people were producing more things. The paradox at the heart of the current moment is that this may no longer be true. Machines are doing more of what economies value. The value keeps accumulating. The people whose labor was the source of that value in the previous century are increasingly beside the point.

12 years ago I was building a platform in West Africa called JOKKO App. It was a design response to a problem that seemed local at the time. Employers could not find the right entry-level talent, graduates could not identify themselves in a labor market in a way that enabled them to grow, over 80% of the African economy being informal, and over 70% of African universities offered no career services towards employability. My assumption was that the identity-in-the-labor-market problem was one of the many systemic challenges found in Africa, made acute by the missing institutional infrastructure. I now suspect it was the early version of a global problem the rest of the world is about to inherit.

The labor share of global GDP is falling. AI is accelerating the mechanism. The knowledge economy is losing its status faster than the categories we use to describe it can keep up. And nobody yet knows what work will mean when intelligence is done by a machine.

1. The shift is real

The International Labour Organization (ILO) reports that labour's share of global GDP fell from 53% in 2014 to 52.4% in 2024. The corresponding rise went to capital. That is one decade of movement. IMF's work documents that labour income shares in advanced economies have declined by roughly 4 percentage points since 1970, and the IMF's own analysis attributes about 50% of that decline to technology. The Bank for International Settlements confirmed the direction in its 2025 Annual Economic Report: automation, not trade, explains most of the manufacturing employment decline of the past decade.

The number that mattered most to me while writing this appeared in an Anthropic Economic Index report published in March 2026. Between August 2025 and March 2026, tasks in the Computer and Mathematical category migrated from Anthropic's consumer chat interface to its direct API. Down 18% on the interface where a human sits in the loop. Up 14% on the interface where the human has handed the task off. Anthropic's own researchers noted, in their own words, that this migration "may signal more imminent transformation of work for the associated jobs". It seems that task by task, category by category, the humans are stepping out.

This is what a labor share shift looks like at the frontier. It is not a strike, not a layoff, not a dramatic quarterly earnings call. It is a quiet migration of tasks from surfaces where humans still touch them to surfaces where they no longer do. The published data will lag reality by 2-4 years, because national accounts do not measure API traffic. So by the time labor share appears in headline statistics as materially different from today's number, the underlying reallocation will already be well established.

There is no economic law that says technology must create jobs. That was an empirical regularity that held while humans were the bottleneck. When they are not the bottleneck, the regularity may not hold.

2. The knowledge economy loses its status

The professions that carried the highest social status through the 20th century (law, medicine, accounting, consulting, journalism, finance, software engineering) share a common feature. Their value came from applied intelligence: reading, synthesizing, arguing, drafting, deciding within a bounded domain. That is precisely what current AI systems do well.

Anthropic's June 2026 survey of about 9,700 Claude users found that early-career workers report AI can do the highest share of their work, and are the most worried about job loss. Workers with 15 years or more of experience report AI can do about 10% less of what they do. The tasks they name as beyond AI's current reach are judgment, contextual awareness, situational reasoning, relational trust, and managing people. The knowledge economy is not being flattened; it is being inverted. The higher-status entry-level positions are the most exposed. The lower-status positions that require standing in a room, laying a cable, fixing a boiler, are for the moment further along the road to safety.

I keep returning to what this means for the humanities layer of education, and I bring back something I noticed when I was living in Sub-Saharan Africa. The West spent decades pushing entry-level talent into coding bootcamps. The premise was that software was where the durable knowledge-economy jobs would be. Those bootcamps have now graduated into the profession most compressed by generative AI. The humanities equivalent, where people would get taught how to think, how to hold values under pressure, how to interpret complexity, was never built. The professions that resemble it (teachers, ethicists, mediators, therapists) are among the least prestigious in the current economy. If the Anthropic finding on judgment and contextual reasoning holds, they are also the most durable. The status ranking is quietly inverting.

The obvious question, and I do not know the answer, is whether the trades that look safe now stay safe. Robotics is behind LLMs still, but it’s moving very fast. Hardware failure modes are still being reduced by design, and a 10-year horizon may look different from a 5-year one. But for now the plumber and the electrician are in a stronger position than the paralegal and the junior analyst, and that is not a temporary reversal. It is a rearrangement of what the labor market values.

3. But labor itself may be dissolving

Here I have to hold two positions in mind and neither is fully comfortable.

The first, which the data supports, is that we are watching a durable shift from labor income to capital income. The mechanism is documented, the direction is clear, the AI acceleration is happening. Anton Korinek, who now sits on Anthropic's Economic Advisory Council and is the most rigorous academic voice on this transition, has been direct: as AI approaches general capability, wages may first stagnate, then decline, then potentially collapse toward the cost of running the machine that does the same work. In his own words, "you could imagine a world in which AI systems and robots can produce everything so cheaply that humans who have to compete with them won't even be able to afford a subsistence income." The tax system, he notes, assumes labor income exists. Much of what we understand as the modern welfare state does the same.

The second position, which the data cannot yet answer, is that the category of "labor" may not survive the transition as we define it currently. If the historical relationship between human effort, income, and productive contribution is being rewritten task by task, the framing of "labor share vs. capital share" may become an artifact of a bookkeeping convention we inherited from the industrial era. This is not the same as saying labor becomes irrelevant. It is saying that what counts as labor may need to be revisited.

Dario Amodei, whose optimism about AI I do not share fully but whose intellectual honesty I do, arrives at exactly this uncomfortable seam in Machines of Loving Grace. His treatment of the meaning question is notably brief and, by his standards, notably vague. He acknowledges human psychological resilience and suggests we will find meaning in ways beyond wage labor. He does not, and cannot possibly, say yet what those ways are. Demis Hassabis has been more blunt. Speaking to TIME in April 2025 he said: "AGI might create abundance, but it won't dispel the incentives for companies and states to amass resources and compete with rivals. We will need a new political philosophy to organize society in this world. Democracy is not a panacea, by any means, and might have to give way to something better."

That is the CEO of Google DeepMind, on record, in a mainstream magazine, saying that the political-economic order we live in may not survive the transition. Whether one agrees with him or not, that sentence is not to be dismissed.

4. Post-labor is arriving unevenly

The strongest thing to say about the demographic map is that the post-labor question does not have one answer, because it does not have one context.

Japan reached a median age of 51 years old in 2026. Europe faces a projected shortage of roughly 44 million workers by 2050 without immigration. In these economies AI adoption is not a policy choice but a productivity necessity. There simply are not enough working-age people to sustain output at current levels. The distribution question, however, becomes more acute precisely because the labor bargain is weaker. When workers are scarce, they have leverage. When productivity is being generated by capital investment in AI infrastructure, that leverage moves elsewhere.

Sub-Saharan Africa sits at the opposite pole. Median age around 19 years old. A multi-decade demographic dividend still in front of the region. The question is not what happens when work disappears; it is how to create productive absorption for a rapidly-growing working-age cohort in economies that (a) have less institutional infrastructure to absorb it and (b) are being asked to leapfrog into an AI-augmented world that may not need the labor they have. This is the problem I was investigating with my JOKKO App venture, transposed to a continent. When I designed the mobile app platform in West Africa, my concern was that human capital potential was being wasted because the labor market could not identify who could do what and collaboration (as the engine for entrepreneurship and job creation) was hindered. That was true when I was building it, and it is true now. What has changed is that even if the identification problem gets solved, the market may not want the labor.

Latin America sits in the middle of the curve. Mexico's median age is around 30 years old. Brazil is around 34 years old. The working-age share of most Latin American countries will peak between 2035 and 2045. The demographic dividend is a real window, closing on a 20-year horizon. The AI transition and the demographic transition will arrive in the same decade. Whether the region can capture the productivity gains before the demographic advantage closes is the central strategic question for allocators and policymakers alike.

Three archetypes. Three different post-labor questions. Any general answer to what post-labor requires from societies has to be qualified by which population pyramid the society sits inside.

5. What a safety net could hold

I will not pretend to have a strong answer to this, and I do not think anyone can yet. But I want to name what is on the table, because the descriptive work is worth doing before the prescriptive work becomes politically possible.

Universal Basic Income (UBI) has been the most-tested instrument in the last decade. Finland ran a nationwide experiment from 2017 to 2018, providing 560 euros a month to 2,000 unemployed people with no conditions attached. Recipients reported significantly better mental health, less stress, and marginal changes in employment. Kenya's long-running GiveDirectly programs show recipients investing in businesses and generating positive spillovers in local economies. The Stockton demonstration in California produced mental-health improvements and, contrary to the standard critique, an increase rather than decrease in full-time employment. Germany's 3-year pilot has now published preliminary findings showing better mental health and greater willingness to pursue education and career changes.

A pattern across pilots is worth naming: the mental-health improvements are consistent and significant, even when employment effects are minimal. That contradicts one strand of critique of basic income (that it damages wellbeing by removing purpose) and confirms another (that it damages the fiscal base of the state that provides it).

Sam Altman's Moore's Law for Everything essay proposes a different instrument entirely: an American Equity Fund, financed by taxing capital and land rather than labor, distributing ownership rather than income to citizens. His napkin math suggests around 13,500 USD / year per adult within a decade. Whether the specific proposal is politically achievable, its central move (taxing capital not labor) is the same move Korinek argues is necessary on economic grounds.

Daniel Susskind, in A World Without Work, proposes what he calls Conditional Basic Income (CBI): an income floor conditional on some form of participation, whether that is training, care work, learning, or community contribution. I am closer to this position than to unconditional UBI, in part because I do not believe an economy of 8 billion people can afford to abandon the participation link entirely, and in part because I think Andrew Oswald's foundational econometric finding remains true: unemployment damages wellbeing more than any other single characteristic, more than divorce or even loss. Income alone does not repair that, but participation might. The classic form of CBI needs to be layered with regional cost-of-living calibration and life-stage adjustments; although the European Union alone, with its enormous heterogeneity of prices and labor market institutions, illustrates how difficult that will be to implement in practice. But the alternative is a system that transfers income without addressing the identity question, and Oswald's data suggests that is not enough.

Descriptively, this is what is on the table. I do not know which of these instruments will be part of the settlement. I suspect the settlement itself will look different in Japan than in Kenya, and different in Kenya than in Colombia or the United States.

6. The open questions

I would rather end with questions than with answers, because I do not believe the answers exist yet and I distrust anyone confident enough to claim otherwise.

• What is a job when intelligence is done by a machine? If the productive contribution that historically defined a job (thinking, deciding, coordinating, producing) is increasingly a machine output, does the category survive? Or does what we call work quietly split into two: paid contribution to a shrinking labor share, and unpaid participation in something else we do not yet have a name for.

• If work has been our primary source of identity for a century, what replaces it? Not "can it be replaced," but what actually does the replacing at social scale. Community, care, creativity, learning are all named. None of them have institutional infrastructure the way wage labor does. Building that infrastructure is a societal project no country has yet undertaken, and it is not clear who has the authority or the resources to start.

• How is the safety net financed when the labor income tax base is eroding? The whole modern welfare state assumes labor income exists to be taxed. If that base shrinks, whatever comes next (CBI, UBI, equity funds, capital taxes, data taxes) needs a different fiscal foundation, and needs it during the transition, not after it. What role does public-private collaboration play when neither governments nor corporates alone can finance or design the transition fairly? Existing conditional cash transfer programs (Bolsa Família, Prospera) offer partial templates. The AI-lab equity fund proposals offer others. Neither is close to being at the scale required.

• How is the humanities layer built? If the coding bootcamps that were supposed to train the entry level are now training people into the roles most compressed by AI, what would a thinking-and-values bootcamp actually look like? Who runs it, who funds it, who accredits it? Is it a school, a workplace, a state institution, or something we have not built yet? My JOKKO instinct then was that identity is a design problem. My instinct now is that the design problem has scaled from a continent to a species.

• And finally, what if the settlement does not arrive? Carlota Perez's framework suggests that every technological revolution has produced a settlement period in which institutions adapt and gains are distributed more broadly. Every past revolution eventually did. But history has cases where the settlement was slow, ugly, and arrived only after a great deal of avoidable damage. There is no law that says the AI transition will produce a good settlement, or any settlement, on a timeline that spares this generation. The paradox that Amodei sketched, in which growth reaches double digits while unemployment does the same, is what a bad settlement would look like if it took a decade to arrive.

I am writing this from the seat of an observer. As a citizen of Europe, and a resident of Latin America. As a former operator in Africa, China, and the United States. As a DeepTech investor watching the convergence of technologies unfold in front of me, in real time, with capital allocation decisions being made every week on assumptions about what work will look like 10 years from now. I do not know what the answers are. But I know that the questions matter, and I know they matter now.

Paul Clastre
Summer 2026

Disclaimer:
Text written by human, research supported by AI, cover picture generated by AI