One of them is a redistribution policy. The other is insurance, and it is the one that will get built.
August 16, 2026
I write my own invoices. Each one has a line with a number of days on it and a rate beside the number. There is no line for what the days produced.
I am a contract developer. I am paid for the days I work, not for the work I finish.
For most of my career those two things were close enough that the difference never mattered. A day of my attention produced roughly a day's worth of software, the contract bought days, and everyone got what they thought they were buying. The unit of payment and the unit of value moved together.
They have come apart. An integration that used to take me most of a week — mapping two systems onto each other, writing the adapters between them — now lands in a morning. The contract has not noticed, because the contract does not measure that. It measures days. So the arithmetic runs like this: I will deliver the entire year's contracted work by the end of October. I will not be paid in November. I will not be paid in December.
Nobody is replacing me. There is no meeting where someone decides I am obsolete, no severance conversation, no headcount reduction that shows up in anyone's quarterly filing. I am not being replaced by artificial intelligence (AI). I am being shortened by it. The work I was hired to do still exists, I am still the person doing it, I am still good at it — and the money now runs out before the year does.
The money does not vanish, of course. Two months of my billing stops being an expense and becomes a margin, and margins are put to their most productive use — which, in this instance, is any use at all that is not me. The productivity gain is real. Somebody books it. Not me.
Every employment statistic in the world will record me as employed for twelve months.
The mechanism is not complicated.
Payment is denominated in time. Value is denominated in output. Any tool that raises output per hour opens a gap between the two, and where that gap lands comes down to a single question: does anyone want to buy the extra output?
If they do, nothing bad happens to anyone. A salaried radiologist who reads more scans than they did three years ago still earns the same salary. The hospital keeps the difference, and if the scans keep coming it hires more radiologists. Nobody in that arrangement ends up worse off than they started. It is what productivity has looked like for two hundred years.
If nobody wants the extra output — if the amount of work is set by the contract rather than by demand — then the person paid for the time absorbs it. I do not have a shrinking market. My client wants what my client wanted before: a defined amount of software this year. Working faster does not raise that number. It just reaches it sooner.
That is the small version. One developer, one contract, one calendar year, fully visible, no forecasting required.
Keep it. It is not about me, and it is not really about AI.
There are two entirely different reasons to want a guaranteed income floor. Almost every argument about the evidence is an argument about the first one. The second does not depend on a single thing the first one fights over.
The first argument is redistribution, and it is for a world that still has work in it.
It goes like this. Wealth in a productive economy is generated collectively and captured unevenly. Where there is plenty of work and plenty of output, a floor spreads both the work and the money the work generates. It reduces inequality directly, in the most boring possible way: by moving money to people who have less of it. Set against most of what governments do, it is administratively cheap and remarkably hard to game.
It fixes something else. Conditional support is priced in the wrong unit too. Unemployment benefits are denominated in not working, and so is the stack of secondary supports attached to them: housing, health coverage, transport, reduced fees. Not working is the quantity the system measures and pays for. So the moment someone works, the payment is withdrawn, and the withdrawal is the real cost of the job.
A universal floor pays you for existing. Nothing is conditional, so nothing is withdrawn, so working always adds to what you have. Whatever else is true about Universal Basic Income (UBI), it is the only proposal that gets the unit right without one new rule, form, or caseworker. Every conditional version has to decide who qualifies. Deciding is where the wrong unit comes back in.
This is not the fundamental case. That is not the same as saying it is a bad case. It is a good one — humane, arithmetically sound, and if it were the only argument for a floor it would still deserve to win.
It is also, in every important respect, an argument about how to divide the proceeds of work. Which means it assumes the proceeds of work.
There is now a serious test of the first argument. Take it in full.
Between 2020 and 2023, a research team ran the largest and most rigorous test of unconditional cash ever conducted in the United States. One thousand low-income people received $1,000 a month for three years, with no conditions of any kind. Two thousand more served as a control group.
The study was run through OpenResearch, an organization funded substantially by Sam Altman. It was published by the National Bureau of Economic Research as Working Paper 32719. The authors are Eva Vivalt, Elizabeth Rhodes, Alexander Bartik, David Broockman, Patrick Krause and Sarah Miller.
Two details about the design matter before any finding does.
The control group was not paid nothing. Controls received $50 a month. That is a real transfer, and it means every measured effect is the difference between a large transfer and a small one, not between a transfer and zero. My reading is that this compresses the measured effects toward the middle; the authors do not say by how much.
And the transfers stopped after three years. Everyone in the study knew they would. Guy Standing, who has spent a working lifetime on this question, has argued on exactly these grounds that the study was not a basic income experiment at all: a payment that is neither universal nor permanent is a different object from a basic income, and people behave differently when they know the money has an end date. He is right that it is a different object. It is still, by a distance, the best evidence anyone has about what happens when you hand poor people money and attach no strings.
So: the finding everyone quoted.
People worked less. The original paper reported a 2.0 percentage point decrease in labor market participation and a reduction of 1.3 to 1.4 hours of work per week, with recipients' partners cutting their hours by a comparable amount. Total individual income, excluding the transfers themselves, fell by about $1,500 a year. The authors' own summary described a moderate labor supply effect that did not appear to be offset by other productive activities.
Concede all of it. It is the study's finding, the authors put it in their own abstract, and an argument that needs it to be false is an argument in trouble. The Heritage Foundation built a case for dependency on it. That case is not built on a misreading; it is built on the paper.
What almost nobody quoted was everything else in the same study.
Recipients increased their spending by $310 a month on average. The largest increases were food, at $67 a month, rent, at $52, and transportation, at $50. Spending on other people rose by $22 a month — a 26 percent increase relative to the control group. Read that last one twice. Give money to people who have very little, and one of the first things they do is give some of it to somebody else.
Then health. Days spent taking painkillers that had not been prescribed to them fell 53 percent across the whole group, and 81 percent among the men. Drinking that got in the way of responsibilities fell 20 percent. The share of people who saw a dentist at all in a year rose 10 percent, and 12 percent among those who had been skipping care they needed when they signed up.
And then the findings that nobody quoted because they help nobody at all.
No measurable effect on physical health — none, across the study's health measures. No impact on the quality of employment, and the authors were blunt that their error bars were tight enough to rule out even small improvements. No significant effect on investment in human capital — schooling, training, credentials. And of all the categories of time freed up by the transfers, the largest single increase was leisure.
That is the honest shape of the evidence. Not a vindication. Not a refutation. A floor bought food, rent, transport, generosity, sobriety and dental appointments; it did not buy health, credentials or better jobs; and it cost some hours of work.
Fortune ran a piece in September 2024 arguing the study had been widely misinterpreted. It had. Be exact about who did the misinterpreting, though, because the answer is close to everybody. Supporters quoted the spending and the painkillers. Opponents quoted the hours. Both sides took the half of a single study that suited them, from a paper whose own conclusion was studiously unexciting.
There is a footnote to this that I find more interesting than the study's supporters generally do. The best-known funder of the most-cited piece of evidence against his own political position is Sam Altman, who supports a basic income. He paid for a result that was immediately weaponized against him, and it was published anyway. I do not need to know anything about his motives for that to be structurally remarkable, and I am not going to speculate about them. Honest publication is penalized in a debate with room for one number. That is a fact about the debate, not about him.
Working Paper 32719 is not a fixed object. It was issued in July 2024 and revised in January 2026, and the National Bureau of Economic Research serves the revision at the same address as the original. It gained an author on the way: Patrick Krause, thanked in the acknowledgments of the first version, is on the byline of the second. If you cited the paper in 2024 and I cite it today, we are both citing it accurately and we do not agree.
Here is what moved.
The 2.0 percentage point decrease in labor market participation is now 4.1 percentage points. The 1.3-to-1.4 hour weekly reduction is now stated as 1 to 2 hours. The $1,500 annual fall in income excluding transfers is now about $1,800.
Two point zero to four point one. The participation effect roughly doubled. The other two figures moved a little. And the revised abstract carries one finding the original did not: subjective well-being rose in the first year and then reverted to control levels.
Revising a working paper is not a scandal; it is what working papers are for. They circulate so that data, methods and measures can be challenged and corrected before anything is final. The abstract does not state why the figures moved, and I am not going to invent a reason.
And the two numbers about to be compared are not the same kind of number. A 4.1 percentage point fall in participation counts people who left the labor market. A 4 to 5 percent work reduction counts work not done. Anyone who slides between them is committing exactly the offense I am complaining about.
With all of that said, one uncomfortable thing is true. For two years, the Heritage Foundation's figure of a 4 to 5 percent work reduction sat well above anything the paper reported. The authors have since revised their own participation estimate to 4.1 percentage points. That does not vindicate Heritage on the merits, because the two constructs still differ. What it does is close the gap that made their figure look invented.
I am on the side of this argument that finds that inconvenient. Complaining that people quote the convenient half of a study, and then skipping the revision that cuts the other way, would be the same offense with better manners. If someone is going to catch me overclaiming, I would rather it were me.
The revision matters most for a different reason. The public debate metabolized the July 2024 number. It has not metabolized the January 2026 one. Both camps are still fighting over a figure the authors have already replaced, because the fight was never about the measurement. It was about whether poor people can be trusted with money. That question is old enough to have been asked in the 1960s, in much the same language, about cash experiments very like these, and it will absorb any number you feed it.
Every one of those findings — the hours, the painkillers, the revision, Heritage, Standing, all of it — is evidence in the first argument's case. All of it is there to settle whether a floor treats poor people well or badly, which is a question about how to share what work produces.
The second argument is not making that case at all.
It is not a redistribution argument. It does not rest on fairness, on inequality, on the dignity of the poor, or on any pilot result ever run. It rests on one claim: that the work will eventually be done without a human checking it, and that when the checking goes, the economic function of most skilled labor goes with it. How long that takes is genuinely unknown. That it is the direction of travel is not.
If that claim is right, the reason for a floor changes completely, while the policy stays the same.
A floor is no longer a way of sharing out what work produces. It is what keeps money in the hands of the people a company needs to sell to, once their hours no longer earn enough to live on.
Redistribution assumes wages coming in and being shared out. It is a quarrel among the people producing the money about who gets what share of it. When the money stops passing through human hands there is no quarrel left to have, and the payment is doing something else: it is keeping the customers solvent. An economy that produces and has no buyers is not an economy. It is a warehouse.
Beneath even that is the older reason, and it is the one I believe moves policy. Take away most people's economic function, leave their income unreplaced, and you are running an experiment history has already run several times. It does not end with a white paper. Large numbers of people with no way to earn, no prospect of earning and nothing to lose are not a social policy problem. They are a stability problem, and the people with the most to lose from instability are the people with the most.
Both arguments get filed under UBI. They are not the same policy.
Redistribution is an argument about who deserves what. Every serious objection to it is a version of that question: who earned this, who is owed that, whether it is fair to the person paying. Insurance asks none of those questions. Nobody argues about whether fire insurance is fair to the houses that did not burn. You buy it because the fire would ruin you and the premium would not, and that holds even if you think your neighbors are careless with matches.
That is the second argument. It is insurance. And notice what it does not require: it does not require the pilots to have shown anything in particular. If the study had found that recipients worked more, the second argument would be untouched. If the revision had halved the effect instead of doubling it, the second argument would be untouched. Whether a thousand people in two American states worked an hour or two less per week is a finding about how people behave when a floor is added to their existing wages. The second argument is not a claim about their behavior. It is a claim about what a floor would be for.
The best evidence anyone has is evidence about the first argument. The second argument was never waiting on it.
The strongest objection to the second argument is not philosophical. It is empirical, it is current, and it is good.
It is this: the displacement is not happening. The fields that were supposed to be emptied out by AI have not emptied out, and radiology is the clearest case. Jack Karsten at Georgetown has put it about as plainly as it can be put: AI is not only not replacing radiologists, it is increasing the amount of work they can do and increasing demand for their services. Cheaper output does not reduce the amount of output demanded. It raises it, and the humans in the loop end up busier than before.
There is a second version of the objection, from the other direction. Firms that cut staff for AI keep quietly hiring people back. The automation turns out to be more expensive, or worse at the awkward cases, than the press release implied. If the substitution were real, the rehiring would not be happening.
Concede it properly, because it is usually conceded badly.
A radiologist on the same salary who reads more scans than they did three years ago has not been harmed. They are paid for the year, they are paid what they were paid before, and the work got easier or at least more productive. If a hospital then hires four more radiologists at the same salary because scans are cheaper to process and doctors order more of them, that is not a disguised loss. That is more people employed at the same pay. It is what productivity growth has looked like every time it has happened, from the power loom to the spreadsheet, and treating it as a hidden injury would mean treating two centuries of rising living standards as one long injury.
The objection cannot be argued around with clever accounting. On the evidence available, it is winning.
It is winning on one condition: that demand for the output expands fast enough to absorb the productivity gain. Where it does, the extra output finds buyers, the hours stay full and the workers keep their salaries. Where it does not, the same productivity gain lands somewhere else entirely — on whoever is paid by the input rather than the output. A contract that buys a fixed amount of work in a year does not order more work because the work got faster. It just ends sooner.
That is the whole disagreement. State it in a form that could be settled rather than debated.
If pay is holding, the first case is what is happening. If radiologists are being employed in greater numbers at the salaries they used to earn, then nothing in the second argument is visible yet, and anyone claiming otherwise is asking you to distrust your own paycheck. The version that would matter is the one where the headcount rises and the pay does not — where the job survives, the title survives, and the terms underneath it are quietly rewritten. That is what the documented rehiring cases show: firms that over-automated, then rebuilt on outsourced and on-demand arrangements rather than restoring the posts they cut. Headcount recovers. The contract underneath it does not.
I am not going to tell you that has happened to radiology, because as far as I can tell it has not.
But look at what the radiologist is doing now.
The machine reads the scan. The radiologist checks it. The workload went up because the machine made them faster, and the job continues because somebody qualified has to be accountable for what the machine produced. Those are two different reasons to employ a radiologist, and only the first is about how fast they work.
The second one has a clock on it.
Every serious deployment of AI into skilled work runs the same way today. The model does the work, a qualified human checks the work, and the human's signature is what makes the output usable by anyone else. That arrangement is productive, and it is why employment is rising in the fields that were supposed to collapse.
Two things keep the human in the chair. The machine is not yet reliable enough to go unchecked, and somebody has to be liable when it is wrong. The first has an expiry date: closing that gap is what the entire industry is funded to do. The second is a rule — European law requires human oversight of high-risk artificial intelligence systems, and professional licensing puts the liability on a named person. Rules like that hold for as long as they are cheap.
So the sequence is not "AI arrives, jobs vanish." There are three steps, and we are in the second.
First, the tool makes the specialist faster. Second, the extra output finds buyers, so the specialist becomes more valuable and more specialists get hired — which is where we are now, and which is why the objection is winning. Third, the checking stops being necessary, and the reason for the employment goes with it.
The third step is the one that decides everything, and it is the one no current data can show you, because it has not happened yet. Employment figures can only report the step we are standing on.
This is also what makes it a question about a floor rather than a question about retraining. The people in that third step are not interchangeable labor. A radiologist is more than a decade of training past school. So is a pathologist, a structural engineer, a tax specialist, a litigator. They spent their twenties, and often their thirties, acquiring one deep competence, on the reasonable understanding that deep competence was the safe side of the automation line. When the checking stops being necessary, that competence does not transfer anywhere. There is no adjacent profession where a decade of reading scans is the qualification.
I cannot tell you when this happens. Nobody can, and you should distrust anyone who hands you a date. The direction is the claim, and the direction is not in dispute among the people building it, because closing the gap that currently requires a human signature is the stated goal.
A floor is the only proposal anyone has that does not require knowing when.
So there are two arguments for UBI, and they are not one argument at different volumes.
The first is the humane one. It says a wealthy society should not have people in it who cannot afford dental care, that conditional support is priced so that taking a job costs you the support, and that handing people money is a startlingly effective way to get them fed and housed and slightly less alone. That argument has three years of hard evidence behind it, some of which cuts against it, and it has been true for as long as there have been poor people in rich countries. It has not won.
The second is the self-interested one. It says that when enough people have no economic function, the arrangement that keeps everyone else's property intact stops being stable, and that a floor is cheaper than the alternative. And the people it is talking about are not an underclass. They are the specialists — the ones who trained for a decade into the work that was supposed to be the safe side of the line. That argument has no pilot data behind it and cannot have any, because you cannot run a randomized controlled trial on a profession losing its reason to exist.
The first argument is the one that deserves to win. The second is the one that will.
That is not a comfortable thing to write, and I have not found a way around it. The serious floors in rich countries — unemployment insurance, pensions, national health systems — were built by people who had looked at what was coming and concluded that the alternative was worse. Not one of them was built because an experiment showed that recipients spent the money responsibly. We already knew that. We knew it from the cash experiments of the 1960s and 1970s, and the 2020s produced a beautifully executed three-year study that told us again, to an audience that had already chosen which half of it to quote.
The evidence was never the obstacle.
I will finish my year's work in October. It is a very small piece of information about a very large question, and I would not build an argument on it. But it does have one property that the study does not: nobody has to decide whether to believe it, and there is no half of it to quote. I will write the invoice myself, and there will still be no line on it for what the days produced. The contract says days. The work says output. The two numbers used to be the same number.
They aren't anymore. That is the whole thing.