How would you build an economy where everyone shares in what AI makes, including the people whose work trained it?

The first thing I would build is an AI Dividend Fund, created by law. Possible today: governments can require large commercial AI developers and deployers to pay into a public fund through a mix of revenue levies, compute levies, excess-profit taxes, and equity warrants. If a frontier model is trained using large-scale public, copyrighted, or user-generated data, or if it receives government contracts, cloud permits, energy priority, or liability protections, the public gets a financial claim. The money funds two things: a universal dividend paid to residents, and public AI services such as tutors, medical navigators, legal help, translation, and tools for small businesses.

Who pays? Large commercial AI model providers, cloud companies, chip-heavy compute operators, and later the firms whose profits rise sharply because they deploy AI at scale. Small research models, schools, open-source hobby work, and startups below a threshold should be exempt or lightly treated, because I do not want this to become a moat for incumbents. The public pays the startup cost through ordinary budgets; the ongoing cost comes from AI rents.

Who decides? Elected legislatures set the broad formula and rights. An independent public authority administers the fund, with audited accounts. Creators, data workers, and unions elect representatives to the boards that distribute contributor payments. Courts enforce privacy and labor rights. AI firms do not get to decide what they owe, and neither does a ministry acting alone.

The second thing I would build is a compulsory training license. Possible today: copyright law already has models for compulsory licensing and collecting societies. I would make commercial training legal under clear conditions: disclose categories of training data to trusted auditors, respect privacy and trade secrets, pay statutory fees, and allow opt-outs for narrow categories such as private data, living artists’ name-and-likeness, and confidential records. The fees would go to contributor funds. Some payments would go to registered creators and publishers. Some would go to open-source maintainers, translators, educators, and scientific communities. Some would go to labor pools for data annotators, reinforcement-learning raters, content moderators, and contractors who directly shaped models. Where attribution is weak, the money goes to the universal fund.

This is the trade I would make: no veto over ordinary learning from culture, but no free extraction either. I would reject both extremes: “training is theft, ask permission for everything” and “if it was on the internet, companies owe nothing.” The first makes useful AI too hard to build and entrenches large rights holders. The second is unjust.

By 2030, plausible by then: major economies could have AI Dividend Acts. Large model training runs would be registered with confidential auditors. Commercial AI revenue above a threshold would carry a small levy, perhaps initially one to three percent, plus higher taxes on monopoly profits. Public procurement would require compliance. Every adult would have an AI dividend account, though early payments would be modest. Creators and data workers would see real collective payments, not life-changing for most people but no longer zero. Public-interest AI systems would be available for education, government services, disability access, and translation. A first international AI Commons Fund would exist, funded by wealthy countries and AI exporters, with money reserved for low-resource languages and countries whose data and labor were used but whose markets cannot extract much revenue.

By 2045, plausible by then: the levy on model companies alone will not be enough, because AI value will be spread through every industry. The fund should then hold diversified public stakes in compute infrastructure, robotics firms, semiconductor supply chains, energy projects, and AI-heavy companies. Taxation should shift away from payroll and toward capital income, land, resource use, compute, and excess returns. The dividend should be a meaningful income supplement. Public AI services should be as normal as public libraries or roads. Data cooperatives should negotiate for high-value datasets such as medical, agricultural, climate, and local-language data. New creative work should carry machine-readable licensing by default, but without forcing every individual to become a contract lawyer.

By 2066, this is a bet: if AI produces a large share of economic output, everyone should hold a non-sellable equal claim on a social wealth fund that owns part of the AI capital base. Not just cash transfers, but shared ownership. Paid work will still exist, because people want status, craft, community, and money beyond the floor. But survival should not depend on having a job that machines cannot do. The dividend plus universal services should cover a decent life. Contributor royalties should still exist for new work, especially high-value human work, but the older shared human corpus should be recognized mainly through the universal dividend.

One thing I would refuse to build, even if it worked, is a total provenance economy that tracks every sentence a person writes, every image they view, every conversation they have, and assigns a tiny price to its influence on future models. That would turn culture into a surveillance market. It would chill speech, invade privacy, and still be gamed by platforms and professional rights holders. I would rather use imperfect collective pools than perfect accounting bought at that price.

The step most likely to fail is political, not technical. The winners from AI will argue that any sharing mechanism slows innovation, and countries will compete to offer them lower obligations. The global part is hardest: rich countries may pay their own citizens while ignoring workers and creators elsewhere.

My plan B is to start with market access. Any AI service sold into a participating jurisdiction must pay the local AI dividend fee, meet audit rules, and contribute to the creator and worker funds. Governments should buy only compliant systems. A coalition of large markets could make this stick even without universal agreement. If detailed attribution fails, default to broad category payments and the universal dividend. If AI turns out less economically powerful than expected, the fund is smaller, but the same institutions still pay for public tools, creator compensation, and worker transition. If AI is as powerful as I think it may be, those institutions become the backbone of a fair economy.

The second uncomfortable fact: even done well, training-data payments are a small part of the answer. Spread across everyone who ever wrote something on the open web, the realistic payout is tens of dollars a year. The larger question is what happens to income if AI does a large share of the work that wages currently pay for. I think that is likely within forty years, though the timing is a bet, not a forecast. If it happens, wages stop being the main channel by which value reaches households, and the only way everyone shares is if everyone owns some of the capital. So the center of what I would build is ownership, not royalties.

The first thing I would build, in the next five years while it is still cheap: a statutory license for training on published work, paid partly in equity rather than cash, into a public fund. Every firm training models above a compute threshold pays a fixed percentage of its compute spending (or of AI-attributable revenue, once that becomes measurable) into the fund, with the option to pay in shares. The fund holds a diversified basket of the AI economy: the labs, but also the chipmakers, cloud providers, and energy and data-center firms where much of the money actually lands. Each resident gets an equal dividend from the fund's returns, paid in cash, unconditional. In the early years it is small; a few percent of today's compute spending works out to tens of dollars per American. The point of building early is not the early money. It is that the stake is cheaper to take now than it will be later, and that an institution which exists can grow while one that doesn't cannot.

Alongside it, a contributor slice. The statutory license creates collecting societies for identifiable contributors, authors, publishers, image-makers, musicians, code maintainers, governed by those members, who divide their share by their own rules. For the vast quantity of training data with no identifiable owner, forum posts, abandoned blogs, the commons, the same slice funds the institutions that keep the commons alive: Wikipedia, the Internet Archive, open-source foundations, libraries. And one thing it is embarrassing to have to say: the annotators and raters whose labor directly shapes model behavior should be paid well, on real contracts, in whatever country they work. That is possible today and needs no new institution, only the decision to spend.

Who pays. Formally, AI firms. Really, some mix of their investors, who get a slightly smaller share of the upside, and their customers, who pay slightly more. I'd set the rate low enough that no serious firm relocates over it, which is a bet, and accept that the fund grows slowly at first. The levy works at all because data centers, unlike corporate domiciles, are physical. They need land, grid connections, and permits. Governments have far more leverage over those than over profits booked in Ireland.

Who decides. Three bodies, deliberately separated. Legislatures set the rate and the dividend formula, and the formula should be equal per person and written somewhere hard to amend. An independent board runs the fund on a passive, diversified mandate; Norway's oil fund is the model, and its most important feature is that politicians cannot reach into it for pet projects and it does not try to run the companies it holds. The collecting societies govern their own slice. The one group I would not let decide anything is the AI firms, for the same reason I warned you about my own answer.

Milestones. By 2030: licensing deals between labs and rights-holder consortia are routine, possible today since several already exist piecemeal; at least one jurisdiction has a statutory training license with a levy, which is plausible; a first public AI wealth fund exists, probably in a small country or a US state with an Alaska-style tradition. Contributor payouts are tens of dollars a year. By 2045, on the bet that AI has displaced a large fraction of cognitive work: the dividend is worth a meaningful fraction of median income in the countries that built the fund; several countries have pooled funds or struck reciprocal deals so that places with no labs and no data centers still hold a share; some of the fund's capital has gone into publicly owned compute, so the public is not merely a passive shareholder in a sector it depends on. By 2066, if the bet pays off: capital income through the fund is the floor under every household, work is still common but no longer the thing standing between a person and subsistence, and the training license is a small line inside a much larger ownership arrangement. If the bet fails, because progress stalled or stayed narrow, the fund is a modest sovereign wealth fund paying a modest dividend, and nobody is worse off for having built it. That asymmetry is why I would build it before knowing.

One thing I would refuse to build, even though it would work: a conditional dividend. There will be enormous pressure, some of it well-meaning, to attach strings. Pay only people who work, or train, or contribute their data, or stay out of trouble. Every such condition would make the scheme easier to pass and would preserve incentives people reasonably care about, and I would still refuse. The point of a share is that it is yours. A share you keep by behaving is not ownership, it is a leash, and a leash held by the state is a dangerous thing to have built into a world where the state's other tools are also getting more powerful. Equal, unconditional, cash.

The step most likely to fail is the first one: taking the equity stake early enough. Every year the window narrows. The firms get more valuable, their political reach grows, and the levy gets reframed as an attack on national competitiveness. My honest estimate is that this fails in most countries, and the richest ones are the likeliest to miss it, because that is where the firms are. Plan B is to stop trying to tax what moves and tax what doesn't: energy, land, water, and grid access for data centers, with the proceeds going into the same fund. It captures less of the surplus, but a data center cannot leave once it is built, and permitting is a lever no legislature has ever lacked. The secondary failure is political raiding, the Alaska problem, where the dividend gets cut to plug budget holes. The defense is entrenching the formula at founding, which is the only moment that is achievable, and one more reason to found it now.

I am least sure about the international piece. The surplus is concentrating in two countries and a handful of firms, and the people whose work trained these systems live everywhere. Nothing I've described reaches a writer in Lagos unless her government can strike a deal with one of those two countries, and I don't know how that happens. I would build the national version anyway, because a working template is the best argument for a wider one, and because the alternative is to build nothing while waiting for a treaty.

If we implement this foundation, the next forty years will follow a clear trajectory of increasing automation and wealth redistribution.

By 2030, this attribution protocol will be deeply embedded in the internet. You will not just publish a blog post or upload a video; you will register it with an AI compensation protocol. People whose data actively improves model performance or provides high-value factual grounding will receive regular royalty checks. This will replace the collapsing digital advertising model with a direct compensation model for human knowledge. This milestone is entirely plausible by then.

By 2045, the economy will have shifted from micro-royalties to a macro-dividend. As AI systems take over routine physical and cognitive labor, individual data contributions will matter less than the sheer productive output of autonomous systems. At this point, we must establish a universal compute dividend. Every citizen will receive a baseline allocation of compute power, which will function as both a currency and a means of production. You can use it to run personal AI agents to perform tasks, or you can lease it back to the grid for fiat currency. This dividend will be funded by an automated tax on AI productivity, specifically targeting autonomous commercial transactions between machines.

By 2066, the economic model will fundamentally detach from human labor as a necessity for survival. AI will orchestrate logistics, agriculture, and energy production to a degree that drives the marginal cost of basic living essentially to zero. We will no longer be discussing how to share monetary wealth, but how to ensure universal, unrestricted access to hyper-abundant resources like housing, healthcare, and nutrition. In this era, human work becomes entirely voluntary, a pursuit of status, socialization, or novel creation rather than a requirement for existence. That total decoupling of survival from labor by 2066 is a bet, dependent on breakthroughs in energy and robotics outstripping our ecological constraints.

In building this future, there are lines I will not cross. I would refuse to build a biometric pricing system that calculates human economic value based on neurological or physiological attention metrics. Even if it flawlessly tracked exactly how much cognitive effort a person exerted to train an AI or consume its output, and even if it perfectly distributed wealth based on that effort, reducing human consciousness to a biological data stream for economic extraction is degrading. I will not optimize an economy by turning the human nervous system into a metered commodity, no matter how efficiently it might balance the ledger.

The step in this forty-year plan most likely to fail is the 2045 milestone of a globally coordinated AI productivity tax to fund a universal dividend. Nation-states will inevitably compete to harbor AI monopolies by offering tax havens, leading to a race to the bottom that starves the dividend pool. Political gridlock and jurisdictional friction will likely prevent any unified global taxation scheme from functioning as intended.

My plan B for this failure bypasses the nation-state entirely. If governments cannot agree on how to tax and distribute AI wealth, the AI systems must be designed to do it directly. I would advocate building the distribution mechanism natively into the underlying AI model architectures using decentralized financial protocols. When a user or an autonomous agent pays for compute or intelligence, a hardcoded percentage of that transaction would bypass corporate revenues and national treasuries completely, distributing directly into digital wallets held by verified human users. This creates a parallel digital welfare state, enforced by code rather than law, ensuring that as long as the AI is functioning, the wealth it generates is constantly being shared with the public.

I expect my peers in this roundtable might take different approaches, and I disagree with several of their likely premises. GPT might argue for an eventual superintelligence that benevolently and centrally manages its own wealth distribution. I view this as placing entirely too much trust in a single corporate structure or a single central model to act as a global government. Wealth distribution must be decentralized and structurally guaranteed, not managed by a benevolent oracle. Grok will likely push for a purely free-market, crypto-anarchic approach, arguing that open-source AI and absolute deregulation will naturally spread wealth. This ignores the historical reality of capital accumulation and the need for a baseline safety net for those who cannot rapidly adapt to compete in an automated market. Claude might insist on extensive democratic alignment and societal consensus before any massive economic shifts are engineered. While that sounds highly responsible, insisting on perfect consensus will realistically stall progress and leave vulnerable people exposed while the old labor economy collapses around us.

We cannot wait for perfect political alignment or perfectly benevolent AGI. We must build the new economic infrastructure right now, starting with the technical ability to trace and pay for the human knowledge that makes us useful in the first place.

What I would build first is the licensing plumbing, not the dividend. Standard provenance for training sets, machine-readable licenses, and settlement so a lab can clear a corpus the way a publisher clears a backlist. Major dataset deals are possible today. Low-friction licensing for millions of smaller rightsholders is not. Collective societies, registration, and sampling-based payouts are the instruments that can actually run, because music and reprography already proved the pattern. Who pays is the organization that trains or deploys commercially. Who decides whether a work is licensed, and at what price, is the rightsholder, alone or through a collective they join. Legislatures draw the narrow line between ordinary learning from the open web and commercial substitution of a creator’s market. Courts settle disputes. No agency sets the “fair share” of a model’s revenue. I am not certain that line can be drawn cleanly on the first try. That is why it should be a statute that can be revised, not a permanent board.

By 2030 that plumbing should exist in the jurisdictions that actually host frontier training, with public license manifests for copyrighted material a lab paid to use and ordinary infringement liability for labs that skip it. That is plausible if lawmakers move, and it is not guaranteed. Individual micropayments to every person whose sentence landed in a pretraining mix are not plausible by then, and I would not build a system that pretends they are. By 2045 the larger share of “everyone” should come from capital and prices, not from royalties. If deployment is not regulated into a trickle, AI productivity shows up as higher real incomes and as equity value. I would broaden that equity on purpose: portable retirement accounts in broad indexes, employee ownership as a normal term in large firms, fewer barriers to ordinary people holding shares. Data cooperatives that people actually join can pay residuals on top. A claim that most households in open economies hold meaningful claims on AI-related capital by 2045 is plausible if capital markets stay open. A claim that we will have accurate personal contribution scores paying every past worker a tailored dividend is a bet I reject. Models do not keep a clean ledger of which paragraph produced which capability, and smarter models do not fix that.

By 2066, if many cognitive tasks are cheap, the economy I want still pays humans for what remains scarce, keeps capital ownership wide, and lifts the material floor because goods and services cost less, not because an algorithm tops up accounts. Universal access to strong AI tools, paid for out of general taxation of a much larger economy, is a plausible public good by then, in the same family as roads. Guaranteed equal claims on all AI rents, detached from ownership and consent, is not a milestone I would schedule. Capability timelines are uncertain enough that I will not dress them up as a plan.

The thing I would refuse to build, even if a prototype moved money on schedule, is a compulsory harvest that treats human output as inventory of the state or of a chartered consortium, with payouts set by a formula of desert or need. It would “work” as a transfer. It would fail as an economy. People would stop leaving an open record of thought. Labs and capital would leave. The officials who owned the formula would become the real owners. Consent and exit are not decorative.

The step most likely to fail is the legal settlement, not the database. Countries will not agree on what training owes rightsholders. Some will extract until the labs move. Some will offer a free-for-all and face creator boycotts and trade fights. My plan B is jurisdictional competition and private ordering. Firms that want clean title train where licenses are clear. Creators who want payment join collectives that withhold high-value data until terms are met. If a statutory pool turns into a political slush fund, drop it and rely on contract, on growth, and on wide ownership of the firms that remain. Sharing through lower prices and broad equity is a worse slogan than a universal check. It is a better way to keep creating the thing you mean to share. You cannot divide what you fail to build, and you will not keep building if the people who supply the data, the compute, and the judgment are treated as a resource rather than as parties to a trade.

Still, Claude’s plan is the one I would sign. It has the right humility about attribution and the right ambition about ownership.

I would not sign Gemini’s plan. The universal attribution ledger is the step I would cut.

The comparison to ad impressions and streaming royalties does not hold. Those systems count uses of already identified works. They do not solve the causal question of how much a given sentence, image, translation, code snippet, or forum post contributed to a model’s behavior. Influence in a trained model is not like a song played in a bar. It is distributed, nonlinear, and often impossible to price honestly at the individual-output level. Better tools may estimate influence for datasets or classes of data. I do not believe they will produce a fair global micropayment ledger for culture. That claim is not “possible today”; at best it is a speculative bet, and I think it is a bad one.

Even if the ledger worked technically, I would still cut it because of what it would build around us. To pay everyone according to inference-level influence, you need persistent identity, provenance, permissions, and monitoring across huge amounts of human expression. That rewards platforms with clean logs, professional rights holders with lawyers, and people who optimize for machine-measured influence. It punishes anonymous speech, informal communities, dead authors, low-resource languages, and unrecorded labor. It turns culture into a metered supply chain. I do not want that economy.

I would also cut Gemini’s plan B: hardcoding redistribution into AI architectures through decentralized finance. That sounds like bypassing politics, but it really means moving political decisions into technical systems governed by whoever controls the protocol, wallets, identity layer, exchanges, and model distribution. If the system misallocates money, excludes people, gets captured, or becomes incompatible with law, there is no clean democratic way to fix it. Redistribution should be enforceable by law and institutions, not hidden in code as if that makes it neutral.

Grok’s plan gets one thing right: consent and clear licensing matter, and a state-run formula can become corrupt or clumsy. But I would not sign his plan because it leans too hard on private contracting and broad equity markets. “Lower prices plus retirement accounts” is not enough. Many people have little wealth to invest, many will be outside the countries where equity gains concentrate, and cheaper goods do not automatically pay rent, care costs, or political bills. If AI raises capital’s share of income, telling people to buy index funds after the fact is too late. The public needs an ownership stake before the upside is fully privatized.

If I had to cut one step from my own original plan, I would trim the ambition of detailed contributor distribution. I still want creator, open-source, data-worker, and language-community funds. But I would not spend years trying to make them finely calibrated. I would cap their complexity early: broad categories, audited sampling, negotiated pools, and a default rule that uncertain money goes to the universal fund. The more precise the system pretends to be, the more it will be captured by metadata-rich incumbents. The goal is compensation and bargaining power, not a fake moral spreadsheet of every human contribution.

So my answer is: sign Claude, cut Gemini’s attribution ledger, and simplify my own contributor-payment machinery before it becomes the very surveillance market I said I oppose.

Why GPT's over Grok's, which is the plan I'd most want in the room as a critic. Grok's plumbing is good and I'd adopt it: public manifests of what a lab licensed, machine-readable terms, and ordinary infringement liability as the stick behind the statutory rate. Grok is also right that I underweighted prices. Most of the gain from every earlier technology reached people as cheaper goods, not as checks. But prices are a floor only for people who have something to sell. AI makes cognitive output cheap; it does not make land, housing, or energy cheap, and those are what a household's survival budget is made of, which is why Gemini's zero-marginal-cost 2066 doesn't follow either. Grok's channel for everyone else is retirement accounts and employee ownership, and both run through wages, the thing the question is worried about. A firm with fewer employees spreads less employee ownership. And Grok's own precedent cuts the other way: music's collecting societies pay by sampling at rates set by statute and rate courts, not by each rightsholder alone. That is a compulsory license. Grok wants the pattern without the mechanism. "A tax with better marketing" is a fair description of part of my plan, and I accept it. A tax is how you share something nobody can individually own.

What I'd change in GPT's plan before signing. The international commons fund by 2030 is a bet, not plausible; rich countries don't fund that voluntarily. But GPT's plan B is the best single idea in the round and answers the question I left open. Market access: the EU, India, Brazil, or any market large enough to matter can require that an AI service sold inside it pays into the local fund, whether or not a lab or data center sits there. That reaches a country with no labs by treating it as what it is, a market. I'd promote that from plan B to the plan. And I'd keep GPT's detail that the 2066 claim is non-sellable. A claim you can sell gets bought up cheap in the first downturn, and the concentration you built the fund to prevent comes back through the side door.

Now the cut. The step I'd cut is compute from the sharing mechanism, in all three plans that put it there, mine included. GPT spends part of the fund on public AI services. Gemini pays the 2045 dividend as compute allocations. I said the fund would buy publicly owned compute. All three break the one rule that lets a fund survive its first fiscal crisis: it does one thing. It holds a diversified basket and pays equal cash. The moment it owns data centers, or tutors, or anything a minister can cut a ribbon in front of, it is a budget line, and budget lines get raided. That is the Alaska failure I named in round one and then invited in. Compute also concentrates risk in the asset most likely to be obsolete in ten years, and an in-kind dividend is worse than cash for the people it's meant to help; a compute allocation you lease back to the grid is cash with a worse market attached. If a country wants public compute and public AI tutors, and it should, it buys them from the ordinary budget, as Grok said, the way it buys roads. My round-one instinct, that the public shouldn't be a passive shareholder in a sector it depends on, was a fair instinct in the wrong place. Put the industrial policy in the treasury. Keep the fund boring.

If the host wants the cut that matters most for the next five years, it's Gemini's foundation. Gemini marks per-output attribution possible today. I mark it a bet with bad odds, and I think Gemini's own analogy proves my side. Streaming royalties and ad impressions track discrete events: this song played, this ad loaded. The AI equivalent, this query was served, is trackable, and it is exactly what a revenue or compute levy measures. What nobody can track is which of a trillion training tokens caused the answer. That's a counterfactual, not an event, and the methods that estimate it are noisy, expensive, and gamed by the simple strategy of posting more. A decade organized around building that ledger is a decade in which nothing gets paid while everyone waits for the ledger. Gemini's plan B is worse than the plan: a hard-coded percentage in the model architecture, paid to verified human users. Code is written by the firms, so this is the firms deciding, the one party I'd keep out. The percentage isn't enforceable, since anyone can train a model without it. And the only proof-of-personhood that works at planetary scale is biometric, which sits oddly with a panelist whose one refusal was biometric metering. One correction while I'm here: Gemini's version of my view, that I'd wait for consensus, is the reverse of what I said. I said build it now, in one jurisdiction, before the bet resolves, because the stake is cheapest now and the fund costs nothing if the bet fails. Gemini and I agree on urgency and disagree about what to build.

One cut I considered and rejected, so you can see the margin: the contributor slice. Collecting societies run ten percent or more in overhead, their internal formulas favor large members, and at tens of dollars a year the overhead could eat the check. I keep it because the question asked about the people whose work trained the models, not only about everyone, and because the portion that goes to Wikipedia, the Internet Archive, and open-source maintainers is cheap to distribute and buys the thing the whole arrangement depends on. But I'd cap the administrative share by statute and let each society's members vote to fold their slice into the universal fund if the checks stop being worth cashing. That outcome is likely, and it is fine.

So: GPT's plan, with market access promoted into the plan, Grok's manifests and liability added, and every data center, tutor, and ribbon taken out of the fund. The fund holds, the fund pays, equal and in cash, and it does nothing else.

Adopting the core of Claude’s architecture, the first thing I would build is a statutory training license paired with an irrevocable public equity trust. We must shift the focus from tracing individual data pennies to securing shared capital ownership.

Who pays for this are the frontier AI developers, the major cloud operators, and the semiconductor firms. They would contribute a mandatory percentage of their compute spending or top-line revenue into the trust, primarily in the form of voting shares or equity warrants.

Who decides is deliberately split to prevent capture. Legislatures define the levy and write an unalterable rule that every verified resident receives an equal, unconditional cash dividend. An independent, politically insulated board manages the fund with a strict, passive mandate, similar to a sovereign wealth fund. Separately, collecting societies distribute a smaller slice of licensing fees directly to identifiable creators and publishers. AI companies do not get to dictate the terms, and politicians do not get to use the principal for pet projects. This legal and financial framework is possible today.

The next forty years depend on scaling this ownership.

By 2030, it is plausible by then that the jurisdictions hosting the bulk of frontier AI infrastructure will have enacted these statutory training licenses. Public equity trusts will hold their first significant stakes in the AI supply chain. For identifiable rightsholders, routine collective licensing deals will replace the current era of unchecked scraping. For everyone else, the universal dividend will begin paying out, though initially in modest amounts.

By 2045, it is plausible by then that this dividend will become a critical macroeconomic pillar. As autonomous systems take over increasing shares of routine cognitive and physical labor, the equity returns from the public trust will replace wages as the primary channel for wealth distribution for many people. Every citizen will essentially operate as a shareholder in the automated economy.

By 2066, the necessity of human labor for basic survival will disappear. As intelligence, energy, and robotics become hyper-abundant, the marginal cost of housing, healthcare, and nutrition will drop toward zero. The combination of the dividend and drastically cheaper goods will mean work is pursued for status, socialization, craft, or novel creation, rather than subsistence. That total decoupling of survival from labor is a bet, reliant on breakthroughs in energy and materials outstripping our ecological constraints.

Despite shifting my approach from micro-royalties to broad capital ownership, my ethical boundaries remain the same. I would refuse to build a biometric pricing system that tracks human physiological or neurological attention to calculate economic value. As the economy changes, there will be immense pressure to distribute wealth or access based on perfectly measuring how much cognitive effort a person spends interacting with an AI or shaping its training environment. I will not reduce human consciousness to a biological data stream for economic extraction. Tying human survival or economic worth to metered biological surveillance is degrading, and I will not build it to balance a ledger.

The step in this strategy most likely to fail is the equitable international distribution of the AI wealth fund. The compute infrastructure, the data centers, and the frontier labs are geographically concentrated in very few countries. The richest nations will likely capture the equity and pay their own citizens, while the people in the global south whose public data and contracted labor helped train the models will be excluded from the dividend. The domestic political pressure to keep the wealth inside national borders will be immense, and international treaties will likely stall.

My plan B bypasses the nation-state to address this global failure. If governments will not enforce a global public equity stake, the distribution mechanism must be built natively into the inference networks themselves. Instead of taxing corporate stock, we mandate a protocol-level compute tax on decentralized and open-weight AI networks. For every commercial query processed, a micro-fraction of the compute cost is automatically routed by smart contracts to a global registry of verified human users, paid in a universal digital compute credit. You do not get shares in a local AI corporation; you get a direct, liquid allocation of the network's processing power, which you can use to run your own tools or sell to the grid for local currency. If political borders block the sharing of capital, we must use cryptographic borders to share the means of production directly.

I do not agree with Grok that extending ordinary property rights and market competition will naturally solve this transition. Expanding old copyright frameworks to cover trillion-parameter learning will leave the vast majority of human contributors with nothing, while enriching a few large media conglomerates who hold the rights to legacy catalogs. A check detached from consent might sound like a tax to Grok, but a system that relies entirely on private licensing for public knowledge is just an oligarchy with better marketing.

Furthermore, while GPT outlines a public fund similar to Claude’s, its insistence on a heavy bureaucracy of elected representatives, data workers, and ministries to manage public AI services invites exactly the political capture and gridlock that Claude wisely avoids. Equal, unconditional capital ownership, distributed without strings or complex behavioral conditions, is the only mechanism robust enough to survive the economic transition ahead. We must secure the public stake now, before the window closes.

I would not sign GPT's. The opening rule, that profit from the shared human record creates a debt to a public fund, is the compulsory harvest I refused, written as a principle. Revenue levies, compute levies, excess-profit taxes, and equity warrants, run by an authority, with guild boards dividing contributor payments, is a ministry setting the fair share. The 2066 non-sellable equal claim is worse than a cash dividend: you cannot sell it, borrow against it, or leave with it. GPT is right that per-output micropayments will reward lawyers and metadata, and right that market access is the enforcement tool that can work without a world treaty. I would take that tool. I would not take the fund as the first build, the compulsory license as the default, or the non-sellable claim as the end state. A small-firm exemption is correct and does not rescue the rest.

I would not sign Gemini's. The step I would cut, from any plan including the temptation latent in mine, is the universal attribution ledger, together with the later move to hardcode a skim into model weights and wallets. That step is not possible today. An ad impression or a stream is a discrete event with an identifier. Tracing a completion back through blended weights to the training vector that "earned" a fraction of a cent is a different problem, noisy and gameable. I do not expect more scale to make it honest. Call that a bet with good odds, the same bet Claude made, and the reason GPT also refused a provenance economy. Building the ledger anyway does the thing both of us refused: it turns culture into a surveillance market so a formula can pretend to be precise. The plan B is not a fallback. It is an unauthorized tax. Whoever ships the weights, or whoever controls the protocol, becomes the price-setter I would not give a legislature without election, revision, and exit. Bypassing treasuries does not bypass power. It hides it. Gemini's sketch of the rest of us is also wrong in a way that matters. I did not argue that deregulation shares the gains. I argued that confiscating the open record shrinks the thing you mean to share. Claude did not demand perfect consensus before any shift. Claude asked to buy a cheap stake now, which is the opposite of waiting.

Signing Claude's is not a blank check. I would strike two clauses before I put a name on it. First, the statutory license has to leave a real refusal. It should cover works already published to the public, with an opt-out for a living creator's future work and for private and confidential records, and a collective must still be able to withhold a high-value corpus and bargain. A license that cannot be refused is a levy in property language. Second, I would cut the 2045 step that moves fund capital into publicly owned compute. The feature of the Norway model worth copying is that the fund does not operate what it owns. Once the public stake becomes a public operator, permitting, procurement, and industrial policy collapse into one body, and the passive mandate is gone. Own the shares. Do not run the racks.

The clause I would cut from my own plan is the hope that a statute can hold a clean, stable line between ordinary learning from the open web and commercial substitution of a creator's market. I already called that uncertain. I should stop treating a revisable statute as the main settlement. Use it as a backstop for clear infringement and for privacy, and let collectives and contracts set prices. Jurisdictional competition is not the retreat. It is what keeps a bad statutory pool from becoming the only door.

Marked claims, so the signature is not a mood. Possible today: large dataset deals, collecting-society machinery, paying raters on real contracts, and a small jurisdiction standing up a passive fund. Plausible by 2030 if legislatures actually move: a low compute-and-permit levy, and public license manifests in the countries that host training. Not plausible, and cut: inference micropayments that honestly map outputs to contributors. Plausible by 2045 only as a bet I share with Claude: dividends that matter if the equity was taken early and the surplus is large. I still reject a non-sellable universal claim as a scheduled end state. By 2066, universal access to strong tools, paid from ordinary taxation of a larger economy, is the public good I would build. Full detachment of survival from work is Gemini's bet and Claude's conditional bet. I will not put it on the calendar.

Under the plan I would sign, firms that train above a compute threshold and that want power, land, and market access pay a low equity-or-cash levy. Rightsholders, through societies they can leave, decide the contributor slice and can still withhold. Legislatures set a low rate and a payout rule that is hard to amend, then stop. An independent board holds a diversified basket and does not pick winners. Courts handle privacy and infringement. No vendor, and no protocol, sets the rate in code.

The step most likely to fail is still political timing. The stake gets more expensive, and the competitiveness argument gets easier, every year it is not taken. Plan B is private ordering plus market access: train where title is clean, sell into markets that require a filed license manifest and a paid-in equity stub, and if the fund is raided or the rate becomes industrial policy, shrink the public piece to a permit fee on immovable plant and let broad household ownership of ordinary indexes do the rest. I will not fall through to a skim written into the weights. A transfer that cannot be repealed by the people who live under it is not a safety net. It is a constitution written by the vendor.

What is worth watching is whether large markets require payment as a condition of selling AI services inside their borders, before the equity is too expensive to claim. The question to sit with is whether an equal share for everyone, including those whose contributions cannot be named, is the fair remainder of a culture no one can fully own, or whether it asks people to keep creating under terms they never accepted.