The Prescription

Abundance is built, not distributed. The scramble for AI dominance is becoming a fight over how to divide the spoils between the most fortunate. That is the wrong fight. The spoils are not the point. What we build the machine to do is the point, and that is still a choice.

One engine, two postures. Only one of them builds.

The first write-up in this series watched well-positioned companies misread AI as a feature. The second argued that structural unemployment is coming and that only its magnitude remains a choice. Both leave the question boards, and policymakers are now circling: what do we owe the people the machine displaces? Start with the experiment that most directly tested the obvious answer.

What the checks bought

Sam Altman saw the reckoning coming before most of his peers, and to his credit, he paid to test the consensus remedy. The study, funded by his money and run by Open Research, gave 1,000 lower-income Americans in Texas and Illinois $1,000 a month for three years, compared with a 2,000-person control group, and measured everything. Findings landed in July 2024, and they deserve a fair reading.

The cash worked exactly where its defenders said it would. Recipients gained agency. They budgeted better, planned further ahead, weathered shocks, and grew more interested in starting businesses. They were 14 percent more likely to pursue education or job training in the final year, and the poorest recipients were 34 percent more likely. They worked about 1.3 fewer hours a week, mostly by choice rather than idleness, and were 10 percent more likely to be actively searching for a better job. Smaller pilots rhyme: full-time employment rose among recipients in Stockton's 500-dollar experiment, well-being and trust rose in Finland's, and the largest basic-income study in the world, across 195 Kenyan villages, found recipients invested and earned more rather than less. Anyone who tells you cash makes people lazy has not read the evidence.

Now read the other half. Three years of guaranteed income produced no measurable improvement in physical health. Gains in stress and mental health faded after the first year. And on the question that decides a working life, study co-author Eva Vivalt was plain: "We don't find any effects on the quality of employment whatsoever." Cash bought breathing room. A transfer can relieve a condition. A transfer cannot change the condition.

The more sophisticated version of the same instinct hands people a stake in productive capital rather than a monthly check. Universal Basic Assets, the framework Marina Gorbis developed at the Institute for the Future, proposes exactly that, and I have argued for it myself in the past. I no longer think the frame holds. The framework has never been tested at scale, and more to the point, the proposal makes the same move as the cash: take the engine's current aim as given, then argue about who gets a share of what it produces. Both proposals start after the defeat. They take the direction of the technology as fixed and bargain over the wreckage.

The boldest transfer on the table

The transfer instinct has just reached a national scale, and the problem behind it is real. On June 1, Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act: a one-time 50 percent tax on the largest AI companies, OpenAI, Anthropic, and xAI among them, paid in shares into a federally managed fund. The proposal would give the fund voting rights, board representation, and a path to public dividends. He modeled it on Norway's sovereign fund and Alaska's oil dividend, and his reasoning is hard to dismiss: the models were trained on the collective output of millions of people, much of it taken without consent or payment, so the public has a claim on what they produce. Even the builders half-conceded the premise. Altman has voiced support for some form of public equity in AI while calling the 50 percent figure too ambitious, and OpenAI has floated a public wealth fund of its own.

Grant the diagnosis in full. Concentration is the danger. A handful of firms capturing the upside, while the gap between the people who own the engine and the people sorted by it widens into something feudal, is exactly the outcome to prevent. But watch what even the boldest transfer accepts. The bill takes the engine's aim as it finds it and argues about who owns the output. A 50 percent public stake in an engine aimed at replacement is still an engine aimed at replacement. The deeper question is never who holds the shares. The deeper question is what the company points its capital and talent at.

The direction is not the wind

The single idea this whole series rests on comes from Daron Acemoglu, who shared the 2024 Nobel in economics, and his co-author Simon Johnson: the direction of technology is not like the direction of the wind, a force of nature beyond human reach. People with power choose it, mostly without noticing they are choosing, and history is a long argument about whether they choose well. Acemoglu separates innovation that makes workers more capable from so-so automation. This kind replaces a person for a modest gain and quietly lowers the value of whatever work remains. The first path raised wages for a century. The second hollows them. In his remarks after the prize, he pointed at the same door this essay walks through: we can "choose a direction for technology that creates more good jobs."

That distinction reframes the entire transfer debate. If displacement were a law of physics, transfers would be the only humane response. Displacement is instead the output of a direction we keep choosing by default, because replacement is the easiest thing to sell to a board reading a cost line. The check and the capital stake are both ways of apologizing for a decision nobody has yet been forced to defend. Aiming the engine at making people more capable and at the problems they can feel, and the question of how to compensate the displaced gets smaller, because there are fewer of them and more worth doing.

Choosing the direction is a two-step process, and most of the argument about AI confuses the two steps. The first act is invention, the frontier model, the protein predictor, the breakthrough. The second is diffusion, the slow, unglamorous work of getting the invention into the hands and workflows of the many. Robert Solow saw the gap in 1987: "You can see the computer age everywhere but in the productivity statistics." Erik Brynjolfsson and his colleagues later gave the gap its modern shape, the productivity J-curve. Transformative technologies demand years of intangible investment in retraining, process redesign, and new business models before their benefits show up anywhere a household can feel them. Invention without diffusion is a press release. Prosperity arrives in the second act.

The proof is already on the table

Aiming the engine at human problems is not a hypothesis. The evidence has been arriving fast, and the strongest of it has been validated since 2024.

Aim it at disease, starting with biology's bottleneck. DeepMind's AlphaFold predicted the structures of roughly 200 million proteins, essentially every protein science has cataloged, and made them available free of charge to more than 2 million researchers in over 190 countries. A fifty-year problem in structural biology, cracked and handed to the world at no cost, and the 2024 Nobel Prize in Chemistry followed. Honesty requires the caveat: predicting structures is not curing disease, and the clinic is the test still ahead. The scale of the attempt is visible, though. Isomorphic Labs, the company built on that foundation, raised 2.1 billion dollars on May 12 of this year, has discovery partnerships with Eli Lilly, Novartis, and Johnson & Johnson, and expects to begin first human trials of AI-designed drugs by the end of 2026. Promise, not proof. But the promise pointed in the right direction.

Aim it at work itself. The strongest field study to date on generative AI and labor, published in the Quarterly Journal of Economics in 2025, followed 5,179 customer-support agents. Access to an AI assistant increased productivity by 14 percent on average and by 34 percent for novices, with almost no effect on the most skilled. Read that twice, because the popular story says AI must widen every gap. In that setting, aimed at augmentation, the measured effect ran the other way: the tool pulled the novice up toward the expert, lifting fastest the same rung the second essay watched disappearing. Augmentation with a number attached, and the opposite of the replacement default.

Aim it at learning. I chair Tech for Africa, so one result from last year sits close to home. In a World Bank randomized trial in Edo State, Nigeria, about 800 secondary students spent six weeks in an after-school program using a GPT-4 tutor with teacher support. Across English, AI knowledge, and digital skills, the students gained 0.31 standard deviations, a result the World Bank team benchmarked at roughly one and a half to two years of ordinary schooling and reported as outperforming about 80 percent of comparable education interventions, with the largest gains going to girls. Six weeks. A short pilot and long-term effects are unproven. The signal still stands: a gap that traditional instruction takes years to close, closed in a month and a half, in a place the frontier labs do not think of as a market.

Disease, work, learning. Three aims, three measurable results, all reported or validated since 2024. None of them happened by redistribution, and none of them happened by accident. Each required someone to decide what the engine was for and then build it that way.

The objection worth taking seriously

The strongest objection is not about whether aiming works but about whether anyone can aim on purpose. Brad DeLong and others argue that nobody can tell in advance which technologies will augment workers and which will replace them, that the gains come from a thousand uncoordinated bets, and that steering the engine mostly slows it down. Some of that is true, and it is a fair check on grand industrial plans.

The objection still proves less than it claims, and the J-curve shows why. Invention and diffusion are different acts. Nobody can dictate the next breakthrough; everyone holding a budget can choose which problems to fund and which proven results to spread. AlphaFold was a deliberate attempt to address a named problem. The Nigeria program was a designed intervention with guardrails, not a happy accident. And a deployment's character is observable after the fact: the QJE researchers could see in the data that the assistant lifted novices the most. Direction does not mean central planning. Direction means refusing to let the default choose for you.

Two poles, one failure

The two models the world is actually running both fail this test, from opposite ends. The United States leads most of the frontier. It largely leaves diffusion to fend for itself, producing extraordinary invention alongside concentration, with the top 1 percent holding 31.7 percent of the country's net worth and labor's share of business output down nearly seven points between 2000 and 2011, still unrecovered. Left alone, that path ends in exactly the gulf Sanders is worried about: a small class that owns the engine and a large one that watches it run. China has chosen diffusion at an industrial scale, installing 295,000 industrial robots in 2024, 54 percent of the world's total, while its output per person sits near 13,800 dollars, roughly a sixth of America's 89,600 in nominal terms and about a third adjusting for purchasing power, with far less freedom in the bargain. One country invents more than it spreads. The other spreads at scale with less to spread.

The synthesis is not a compromise between them. A third position takes the American gift for invention, joins it to the discipline of diffusion China is demonstrating, and points to problems whose solving means abundance for the people not in the room when the engine gets built. Health that gets cheaper and better. Learning that reaches the student a tutor never could. The cost of a life, brought down by capability rather than by subsidy. A lab that aims to use even part of its capability that way does more for the people Sanders is trying to protect than any equity stake in a machine pointed the other way.

One bet, aimed

Now bring the argument down from nations to the room where you sit. The direction of technology is not set in a ministry. The aim is to set one budget line at a time, by people with titles like yours and mine. I studied computer science in the early 1980s and have spent four decades watching technologies that could have gone either way go the way the powerful pointed them. The capability is always real, the direction is always contested, and the people insisting the market will sort it out are usually the ones who benefit from the default.

The first essay in this series asked boards one question: which curve is your AI on? This essay adds the second: what is your new-curve bet aimed at? Here is the prescription in one move. Take one AI investment, this year, whose explicit success metric is a human outcome rather than a cost line. A health outcome your patients would notice. A learning curve your newest hires climb in months instead of years. A cost that your customers stop paying. Then fund its diffusion, not just its pilot: the retraining, the workflow redesign, the boring second act where the J-curve pays out. And measure it the way the economists measured the support desk: track who gets better, and watch the entry rung. If your newest people are not improving fastest, you have bought automation and named it augmentation.

The Borlaug standard

One man has already run the full prescription before AI gave us the excuse to debate it. Norman Borlaug spent his career breeding semi-dwarf, disease-resistant wheat at a research station in Mexico, then did the diffusion work himself, carrying seed and stubborn persuasion into India and Pakistan against bureaucratic resistance on every side. Harvests roughly doubled. Famines predicted as certainties never arrived. By the common estimate, first offered by the journalist Gregg Easterbrook, the work saved as many as a billion lives, and Borlaug received the 1970 Nobel Peace Prize because real abundance turned out to be a foundation for peace. The entire program was funded not by a superpower but by the Rockefeller and Ford foundations, philanthropy aiming an engine at hunger on purpose. The Green Revolution carried real costs, from water stress to fertilizer dependence, and an honest account keeps them in view. It remains the rare example of invention joined to diffusion at a civilizational scale. Borlaug did not distribute grain. He built the capacity to grow it, then walked that capacity into the fields that needed it most.

The jet age makes the same point from the commercial side of the ledger. When the Boeing 707 entered service in 1958, flying was a luxury for the few. The innovation that mattered was less the speed than the economics: an airline could carry far more people at a far lower cost per seat, and over the following decades, helped by deregulation and fare competition, flying moved from elite luxury to mass-market transportation. The industry's largest rewards arrived precisely because the invention changed the economics for the many. An innovation pays its biggest dividend when it reaches everyone, not when it stays priced for the people who already have access. Boeing was chasing commercial advantage, not a public mission, which sharpens the point: the biggest prize went to the invention that brought everyone along. The labs deciding what to build next are standing at the same fork.

So the trilogy ends where it has been pointing all along. Remember the graduates from the first essay, the ones who booed the optimists at commencement. They were never afraid of the machine. They were afraid that the people holding it would point it at them, and the fear is reasonable. The answer to that fear is not a check that concedes the fear was right. The incumbent's calm was a choice. The magnitude of the reckoning is a choice. And abundance is the widest choice of the three, made daily in funding decisions, not once in a manifesto, by people who decide whether the most powerful engine of our era gets aimed at the dark or at the light. We do not have to distribute our way out of scarcity. We have to build our way into abundance, and then have the discipline to make sure what we build arrives.

References

OpenResearch, Unconditional Cash Study findings (July 2024); NBER Working Papers w32711 (health) and w32719 (employment); Eva Vivalt co-author remarks via CBS News (July 2024); University of Michigan Ross School of Business summary (2024).

Stockton Economic Empowerment Demonstration, West and Castro evaluation (2021); Kela, results of the Finnish basic income experiment (May 2020); GiveDirectly, early findings from the Kenya long-term UBI study (December 2023).

Marina Gorbis and the Institute for the Future, "Universal Basic Assets" (April 2017); noted here as carrying no large-scale empirical track record.

Bernie Sanders, "The Public Should Own Half of the Big A.I. Companies," op-ed (June 2026), and the American AI Sovereign Wealth Fund Act, introduced in the United States Senate June 1, 2026 (one-time 50 percent tax on the stock of the largest AI companies, paid in shares to a federally managed sovereign wealth fund; modeled on Norway's sovereign fund and the Alaska Permanent Fund). Fortune (June 3, 2026) and Fox Business on the proposal and on Sam Altman's response supporting public equity while calling the 50 percent figure too ambitious.

Daron Acemoglu and Simon Johnson, Power and Progress (PublicAffairs, 2023), on the direction of technology and so-so automation; Daron Acemoglu, remarks following the 2024 Sveriges Riksbank Prize in Economic Sciences (NobelPrize.org, October 2024). The prize, shared with Simon Johnson and James A. Robinson, recognized studies of how institutions affect prosperity.

Robert Solow, New York Times Book Review (July 12, 1987). Erik Brynjolfsson, Daniel Rock, and Chad Syverson, "The Productivity J-Curve," American Economic Journal: Macroeconomics 13(1) (2021).

John Jumper et al., Nature 596 (2021) and Josh Abramson et al., Nature 630 (2024) on AlphaFold; Nobel Prize in Chemistry 2024 (Demis Hassabis and John Jumper, with David Baker); Isomorphic Labs Series B announcement (May 12, 2026); Isomorphic Labs partnership announcements with Eli Lilly and Novartis (January 2024) and Johnson & Johnson (2026); Bloomberg coverage of the clinical-trial timeline.

Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, "Generative AI at Work," Quarterly Journal of Economics 140(2): 889-942 (May 2025).

World Bank, "From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria," Policy Research Working Paper 11125 (De Simone, Barron Rodriguez et al., May 2025), and World Bank Education blog commentary on the Edo State randomized trial.

International Monetary Fund, World Economic Outlook (October 2025), GDP per capita series; Federal Reserve Distributional Financial Accounts (Q3 2025) on the top 1 percent share of net worth; U.S. Bureau of Labor Statistics on the nonfarm business labor share (series PRS85006173); International Federation of Robotics, World Robotics 2025.

Brad DeLong and Noah Smith, Hexapodia podcast (July 2023), on the impossibility of identifying labor-augmenting technologies in advance.

Gregg Easterbrook, "Forgotten Benefactor of Humanity," The Atlantic (1997); the World Food Prize Foundation on Norman Borlaug; Nobel Peace Prize 1970. The one-billion-lives figure is a widely cited estimate, not a precise count.

Boeing 707 entry into commercial service (Pan American World Airways, New York to Paris, October 1958); Smithsonian National Air and Space Museum, "The Jet Age," on the expansion of air travel to a broad public.