I ended the last piece saying the gap between the work disappearing and the new work arriving needs rails, and that I’d make that case properly next time.
This is next time. And the argument is simpler than I expected when I started writing it.
Universal basic income stopped being a fringe position about eighteen months ago. It is now the default assumption underneath almost every serious proposal for what happens to people as AI absorbs their work — including proposals from people who would never use the term. What nobody has done, in any of them, is answer the question that determines whether it works: who exactly gets paid, on what basis, and how do you decide that at national scale without the whole thing being gamed or captured?
We are going to need that answer sooner than the people arguing about it believe.
Start with the accounting rather than the politics.
AI is producing an enormous amount of value and returning it to a very narrow set of owners. That’s not a moral claim. The gain accrues to whoever owns the model, the compute, and the deployment. The cost lands on the people whose labor was the input being replaced. Those two groups barely overlap, and the gap between them is widening faster than any mechanism we have for closing it.
That’s the whole problem in one sentence, and every proposal on the table is an attempt to repair it.
What makes this different from previous automation waves isn’t the displacement — we’ve absorbed displacement before. It’s the speed, and the fact that this time the productivity gain is almost entirely capturable by the owner. A spreadsheet made an accountant more valuable. An agent makes an accountant optional, and the surplus goes to whoever runs the agent.
Every prior transition gave the labor market decades to absorb the shock, because every prior transition had something physical to install — a machine in a lobby, a PC on a desk, a line run to a building. Agents have none of that. Distribution is an API call and propagation is measured in months.
So the question stops being whether people need support and becomes how fast you can deliver it.
Line up the serious proposals and they disagree about almost everything except the diagnosis. They differ mostly on where the money comes from — and, importantly, they’re all nearly silent on where it goes.
The straight UBI position, and the one with the longest pedigree. Andrew Yang’s Freedom Dividend put a number on the American version — $1,000 a month, every adult, unconditional, no means testing — and made it a national platform in 2020, before most people had used a language model.
The argument for unconditionality is more practical than most critics assume. Means-testing costs a great deal to administer, catches fraud badly, and humiliates the people it processes. An unconditional floor eliminates the entire eligibility apparatus, which is both cheaper and faster. Nobody applies. Nobody gets determined. Nobody appeals.
The objection is equally practical. A universal unconditional transfer is the largest fraud target ever constructed, because every dollar is a bounty on convincingly faking a person. That was a manageable problem when faking a person required documents. It’s a different problem when it requires a model.
Some version of this ends up as the base layer of whatever we build. It’s the only design that moves fast enough.
Elon Musk’s Universal High Income is the most optimistic position on the board: AI and robotics make goods and services so cheap that the binding constraint stops being distribution and becomes desire. Work becomes optional rather than necessary. Not a basic income — a high one.
I don’t think it’s wrong so much as it’s a claim about a destination being used as an argument about a transition. If the abundance arrives, the question dissolves. If it arrives in 2045 and the displacement arrives in 2028, everything that matters happens in the seventeen years in between, and this position has nothing to say about them.
Eventually is doing an enormous amount of work in that sentence, and eventually is where people live.
Bernie Sanders’ American AI Sovereign Wealth Fund Act, introduced June 18, is the most mechanically specific proposal anyone has put on paper. A one-time 50% tax on the stock of the largest AI companies, creating roughly $7 trillion in public ownership, managed by an Independent Commission for Democratic AI — seven members, presidentially nominated and Senate-confirmed — holding voting shares. A 5% annual dividend would pay out more than $1,000 per American per year. Companies running both AI and non-AI businesses would have to separate them.
Whatever you think of it, notice what it does that nothing else does: it names an asset, a transfer mechanism, and a payout figure.
The mechanical problem is timing. You’re asking for the equity at the exact moment it’s most valuable and most politically exposed, right as these companies list. The behavioral response is entirely predictable — restructure, reincorporate, delay, relocate.
There’s a softer family of proposals on the same spectrum that I find more durable: tax the flow rather than the stock. Bill Gates proposed exactly that — a levy on AI tokens and robots, on the grounds that “the tax system nudges you toward replacing people with machines.” He called the required change “greater than any in my lifetime.” A per-inference levy is administratively simple, requires valuing nothing, scales automatically with adoption, and grows precisely as displacement worsens. That last property is the one you want.
Gates paired it with something stranger — “Human Reserved” occupations kept off-limits to AI even once machines can do them, child care and jury service among his examples, possibly up to 40% of jobs initially, though he concedes that’s extreme. I don’t think protected categories work; they become the thing everyone routes around. But it matters enormously who is saying it. When the man who spent a decade making the abundance argument starts drafting protectionism, that’s an update about transition speed, not about the destination.
Sam Altman’s Universal Basic Compute: rather than dollars, everyone receives a slice of the model’s compute, which they can use, sell, or donate.
This is the most intellectually interesting proposal on the list, on the theory that the scarce good of the future isn’t money — it’s intelligence, and handing someone a productive capability beats handing them its cash equivalent.
It’s also the most institutionally unsettling. Mechanically, it makes an AI company the distributor of a public entitlement: metering out a resource people are entitled to as citizens, setting the allocation, and running the identity check that determines who counts. “The company that displaced the jobs also administers the compensation” is not a stable arrangement.
The quietest of the five, and the one already running.
Companies are increasingly treating their own tokens and compute credits as quasi-currency — compensation, incentives, distributions to the communities that support a product. Anyone who was in crypto between 2017 and 2022 has watched this experiment run at length, with published results. Gitcoin distributed more than $50 million in matching funds through quadratic funding. Optimism ran five rounds of retroactive public goods funding and published a post-mortem admitting that large rounds devolve into popularity contests.
That’s the honest baseline for anyone designing a distribution program. It’s also the only body of live operating experience we have, and it’s sitting in governance forums no policy staffer has ever opened.
Ask every one of these proposals the same thing. On the fifteenth of next month, who specifically is entitled to this, and on what basis?
Underneath all of it sits a problem no welfare state has ever had to solve. Every distribution system ever built assumed the hard identity question was which human you are. The new question is whether you are one. When 53% of observed internet traffic in 2025 was automated — with agents broken out as their own category for the first time — a transfer system that can’t tell a person from a script isn’t a transfer system. It’s a faucet.
And there’s a second fork underneath that one: do you pay people for being, or for doing?
Pay for being and you get the fastest, cleanest, most fraud-exposed design. Pay for doing and you avoid the pure sybil bounty but commit to building a scoring apparatus for human worth — which at national scale means a state grading citizens on usefulness.
The best available answer is to separate them completely. An unconditional floor tied only to verified personhood, with no evaluation of anybody. Then, entirely separate from it, a narrow layer that funds contribution. Trying to make one system do both is how you attach a social score to subsistence.
We rehearsed the failure mode in 2020. The CARES Act payments went out under near-ideal conditions: overwhelming political consensus, a functioning IRS holding tax records for most of the population, a one-time transfer, simple eligibility, and no meaningful opposition to speed. The GAO found roughly $1.4 billion went to about 1.1 million dead people. People without direct deposit waited months, and the longest waits landed disproportionately on the unbanked and the recently moved.
That was easy mode. Now make it recurring, make eligibility contested, remove the consensus, and run it in an environment where a meaningful share of the entities requesting payment aren’t people.
The risk isn’t that people never get paid. It’s that the latency becomes the crisis. The interval between a policy passing and money arriving is where evictions happen, where radicalization happens, where the story about whether the system works for you gets written.
If displacement compresses into quarters while the response arrives on legislative time, the failure isn’t economic — it’s civilizational. Societies survive being poor. What they don’t reliably survive is a large cohort of capable people who conclude, correctly, that the system has no role for them and no intention of finding one.
There’s a subtler cost too. A floor stops people falling; it doesn’t build anything. What displacement takes first isn’t income — it’s the sense that your work matters to somebody. Meanwhile there’s an enormous amount of work society obviously needs and markets refuse to price: care work, elder support, local infrastructure, teaching, environmental restoration, open-source maintenance, oversight of the agents now doing everything else. None of it has an employer because none of it captures value. Funding that work is the second half of the answer, and almost nobody is designing it.
When I said last time that nobody is building the rails, I was wrong. The payment layer is roughly two-thirds built.
Proof of personhood is live at scale. World runs Orb-based verification, a wallet, an L2 that prioritizes verified humans, and fiat on-ramps, across dozens of countries. Object to iris biometrics if you like — but the architectural claim has been demonstrated: you can establish that a recipient is a unique living human and pay them repeatedly with no caseworker in the loop.
Settlement got cheap and programmable. Stablecoins clear for cents. And the agentic payments buildout of the last eighteen months produced exactly the primitive a recurring entitlement needs: scoped, time-bounded, amount-limited, auditable payment authority.
Nobody is building eligibility. Governments have the legitimacy and systems that are administrative rather than programmatic — designed around applications and caseworkers, which is precisely the architecture that collapses at scale. AI companies solving it from their side puts the firm that displaced the jobs in charge of the compensation. Protocols have the mechanism design and none of the standing, and standing is not optional when the payment is somebody’s rent.
That’s the actual gap. Not the pipe — the rules, the scoring, the appeals, the fraud budget, the published metrics. It’s a governance problem, and governance problems don’t have vendors.
My honest read on timing: the political trigger is a sustained unemployment print above 10% with a visible AI attribution, and the politics move within one cycle of that number rather than in anticipation of it. Legislatures don’t pre-build safety nets; they respond to visible pain.
The useful things to do are the unglamorous ones. Settle the personhood layer while it’s still a technical question rather than a political one, and make sure there are several interoperable providers rather than one, because routing human subsistence through a single private verifier is counterparty risk we’d never accept in a utility. Decide whether the floor is unconditional before there’s a constituency for conditioning it. Build the contribution layer as something narrow, retroactive and separate.
UBI is going to happen in some form. The argument about what to call it — basic, high, a dividend, a citizen’s stake, an allocation of compute — is the argument people are having instead of building the thing every version of it requires.
The pipes are nearly ready. The question of who stands at the other end of them, and what we’re asking them to do, has barely been asked. We have less time to answer it than anyone currently believes.