Solidarity AI 2026
Who pays for AI? Notes for a reverse panel in Bangkok
2026-11-11
On this page
I am attending Solidarity AI, the Platform Cooperativism Consortium’s conference at Chulalongkorn University in Bangkok.
1 My session is “Who Pays for AI?”
I am on Concurrent Session 7: Who Pays for AI? Supply Chains, Data, and Design. It is a reverse panel: each speaker opens with a short statement, and the audience’s questions drive the rest.
My co-panellists:
- Julian Tait
- Eight years at Open Data Manchester, of which he is co-founder and CEO.
- Yiyan Wang
- A researcher in 3D AI and computer graphics at the University of Tokyo, presenting I2V-FOCAL, which builds editable 3D simulations of intersections from existing roadside cameras so that cities, rather than autonomous-vehicle manufacturers, can test vehicles and investigate crashes.
- Me
- Gradual disempowerment: how step-by-step automation could erode human influence over the economy, the state and culture.
These are notes for me and for anyone in the room who wants the links afterwards. I am also one of the session’s community-notes reporters, so the notes from the session itself will land at the bottom.
1.1 My 2 baht
The problem of who pays for AI, I would like to interpret in terms of “what price can be demanded?” That is, I want to think about automation, and the data needed to automate human jobs, in terms of how much value can be extracted by those who are about to be automated, by those who generate the data.
Two texts from the 1950s I think do a great job at summarising this problem:
Vonnegut (1952): in the far future of the 2000s, Rudy Hertz, master machinist, has his skills recorded to tape, to enable the factory to function without needing any more workers like him. A few years later, chaos ensue, as redundant workers resort to violence seeking leverage in a society that no longer needs them.
Another one, from the next year: Shapley (1953) is the core concept in the field we call “cooperative game theory”, the study of how to fairly allocate the value created by coalitions of people choosing to work together. A common feature of With regard to the AI machinery, we are all, individually, dummy players: we add nothing.
Shapley values are forward looking: they consider how much we could fairly demand for our contributions before they are made. They are grossly over-optimistic estimates of our leverage when the contributions have already been made and the value has been extracted already.
Something something cooperatives, and government, and data unions, and above all ownership. Something something resource curse.
1.1.1 Data is a stock, labour is a flow
Labour unions bargain over a flow: work has to be done again every day, so withholding it hurts. A training corpus is a stock: once it has been ingested, withholding it hurts nobody. So the bargaining power of a data cooperative lives in the flow — data that is fresh, local, situated, or private, and therefore has to keep being collected. Julian’s uncollected data from people in vulnerable circumstances, and Yiyan’s continuous footage of a particular intersection, are both flow data. Archives of past writing are not. If this is right, data cooperatives should bargain over ongoing access rather than one-off licensing of what they already hold. See Data ownership in the AI era.
1.1.2 Your hundred bytes
The per-capita arithmetic from the opener is in which are your 100 bytes?. Any scheme that values an individual’s contribution by its marginal effect on a giant model will find that the contribution is nearly zero, because almost everything one person writes is nearly substitutable by what someone else wrote. Shapley values formalize the fair-share idea, and they do not rescue the individual either: the value appears when we evaluate coalitions, not people. So “pay me for my data” is a weak demand, and “bargain with us as a bloc” is a strong one.
1.1.3 A small collective can steer a model
The result from the opener is Hardt et al. (2024)’s algorithmic collective action: a group of users who coordinate how they modify their data can push a deployed learning algorithm where they want it, and the group can be a small fraction of the user base. The reciprocal quantity is the platform’s performative power, its ability to steer users back; monopoly increases it and outside options erode it. Notes at algorithmic collective action. This is a technical argument that pro-cooperative people seem not to have heard much about, and it cuts in their favour.
1.1.4 The commons was the subsidy
Much of what trained the models was a free by-product: someone fixed a bug and wrote it up, someone edited a Wikipedia page. Acemoglu, Kong, and Ozdaglar (2026) model this. When AI supplies the answer without the learning, people stop producing the by-product, and past a threshold of AI accuracy the human-generated stock of general knowledge can collapse altogether. Two results from that model are useful in this room. First, the welfare-maximizing accuracy of the AI is finite: there is a point past which a better recommender makes everyone worse off, because nobody is left to feed it. Second, better pooling and sharing of human knowledge raises welfare unambiguously and makes collapse harder to fall into. Pooling is exactly what cooperatives do. Stack Overflow and Wikipedia numbers and the rest are at Knowledge collapse and the epistemic commons.
1.1.5 Open data and enclosure
When the main reader of open data was a citizen or a researcher, publishing openly was a solidarity strategy. Now a large share of the readers are training crawlers, and publishing openly is, among other things, a gift to whoever owns the biggest model. I expect the incentives to push towards secrecy: less open publication, more private data, more knowledge held back as a competitive moat. That is enclosure, and a cooperative that simply locks its data up reproduces it. The interesting space is between the two, a governed commons in Ostrom’s sense, which I gesture at in the public sphere and its business models. This is the question I most want to put to Julian, given that his organization is literally called Open Data Manchester.
1.1.6 Complexity favours the bigger computer
A popular hope is that AI agents will negotiate on our behalf, driving transaction costs to zero so that private bargains replace regulation. My Hyperscaling Coase essay argues that cheap negotiation does not level the field. Working out what each party’s agreement is worth — how pivotal they are — is computationally hard, and the side with the bigger compute budget estimates it better. A firm that knows which members of a bargaining club are underpricing themselves knocks on those doors first. Personalized offers erode posted prices, and posted prices were what let outsiders tell whether a deal was any good. For a cooperative bargaining with a lab, that suggests insisting on simple, uniform, public terms, because granularity is where the bigger computer wins.
1.1.7 A cooperative is a constitutional dependency on humans
The gradual-disempowerment argument (Kulveit et al. 2025) is that institutions served humans because they depended on human participation, and that the dependence is contingent. One member, one vote makes it non-contingent, at least on paper. Humans are currently also the only legal persons, which is a political advantage I discuss in an orderly retreat from economic relevance. The provocation for the room: would your cooperative let members vote through their AI agents? If yes, the dependency on humans is one delegation away from being nominal. If no, how would we enforce that?
1.1.8 Paid to train the replacement
Data labelling and preference-rating work, much of it in the Global Majority, is the human participation that AI development still needs. It is also the participation most directly being automated: the people paid to train AI are already outsourcing their work to AI, which in turn makes the models worse (Shumailov et al. 2023). “Who pays?” has an ugly answer here: the workers whose task is to make themselves unnecessary.
1.2 The game, written down
Most of the framings above have a cooperative-game-theory version, and writing it down helps me see which claims are doing the work. There are two kinds of player: humans \(H\), who supply labour, data and attention, and an owner \(K\) of capital and models. A cooperative game assigns each coalition \(S\) a worth \(v(S)\), which is what that group could make on its own. Automation shows up as the owner’s go-it-alone worth creeping up on everyone’s together: write \(v(\{K\}) = a\,v(N)\), where \(N\) is everyone and \(a\) rises towards 1 as machines substitute for people.1
1 These are full-information models: every player knows what every coalition is worth, and coordinating costs nothing. Neither holds. Take the deals in which platforms licensed their users’ posts to AI labs. On the model, the users who wrote the posts are a coalition with obvious worth and should have demanded a share. In practice they were not at the table, many learned of the deals from the news, and none knew what their posts were worth to the buyer, who could estimate that far better than they could. The company holding the terms of service did the bargaining. So the models say who would have leverage if they knew their worth and could act together. Supplying those two conditions, information and coordination, is most of what organizing is; the gap between the model and the world is where a union organizer or a cooperative earns its keep.
1.2.1 The dummy player
Shapley (1953) proposed a rule for splitting a coalition’s winnings: pay each player their average marginal contribution, averaged over every order in which the coalition could have assembled. One of the axioms that pins the rule down is the dummy player axiom: a player who adds nothing to any coalition gets nothing.2 That sounds fair, and it is also gradual disempowerment in one line. As \(a \to 1\), humans become dummy players and the fair rule pays them accordingly, without anyone having stolen anything. Cooperative principles already reject contribution-based fairness — one member, one vote, limited return on capital — so in this vocabulary a cooperative is an anti-Shapley allocation rule, adopted on purpose.
2 Shapley’s paper came out in 1953, the year after Player Piano.
1.2.2 Who is in the core
A division of the winnings is in the core if no coalition could do better by walking out. The owner can walk out with \(a\,v(N)\), so any stable division leaves humans at most \((1-a)\,v(N)\). A cooperative’s egalitarian split survives only while the owner cannot secede and do as well alone, which is to say, only while the members hold something the owner cannot get elsewhere: flow data, the street itself, legal personhood. I called such things determinants of power when thinking about bargaining with AI successors.
1.2.3 Threat points
In Nash bargaining the surplus splits according to what each side gets if talks fail. Before training, a withheld corpus is a cost to the lab. After training, the lab’s fallback already includes the corpus, and the data-holders’ threat point is zero. That is the stock-versus-flow framing in one line.
1.2.4 Rudy Hertz’s tape
In Vonnegut (1952) the master machinist Rudy Hertz has his movements recorded on tape, and the tape drives the factory’s lathes ever after. Rudy is proud of it, and then he is redundant. Noble (1984) gives the history behind the fiction. Postwar machine-tool automation had two candidate designs: record-playback, which captured a machinist’s movements while they machined a part, and numerical control, which moved the programming into an office. Noble argues that management and the Air Force chose numerical control for control of the shop floor at least as much as for efficiency. The two accounts fit together: record-playback leaves machinists needed for as long as the parts keep changing, and Rudy’s tape runs forever because his part never changes. Changing parts are flow data; Rudy’s part is the archive.
The present-day analogue is retrieval versus training. Retrieval at query time keeps data a flow that can be metered, attributed and withdrawn; training makes it a stock in the weights. A cooperative does not get to choose the frontier labs’ architecture, any more than the machinists chose GE’s. It does choose the design of its own tools, and the endpoint through which its flow data is served. Even there, a lab that retrieves our data also sees what it retrieved and can train on it next time, so a “no training on retrieved content” clause is a promise, which brings us to Powell below.
1.2.5 Minimum winning coalitions
Riker (1962)’s size principle says that when what matters is winning, rational coalition-builders form minimum winning coalitions, just big enough to win, because every extra member dilutes the spoils. Baron and Ferejohn (1989) add that whoever proposes the split keeps most of the surplus; in platform terms, whoever writes the terms of service. The dual is divide-and-rule: Acemoglu, Robinson, and Verdier (2004) model a ruler who stays in power by paying selected groups so that no two of them have aligned interests against her. The licensing deals AI labs have struck with a handful of large publishers and platforms fit this pattern: pay the pivotal blocs, leave everyone else out, keep the rest fragmented. Per-person data dividends, however well meant, keep the partition fragmented too. Hardt et al. (2024)’s collective action is the same game seen from inside the bloc. More at coalition games.
1.2.6 A rising power cannot promise
Fearon (1995) asked why rational states fight instead of bargaining. One answer, which Powell (2006) develops, is a commitment problem: when power shifts fast, the rising side cannot credibly promise to honour today’s deal once it no longer needs to. Allison (2017) popularized the same structure as the Thucydides trap, with US–China rivalry as the case. Historians dispute his tally of historical cases, but the mechanism is Powell’s and does not depend on the tally. Windfall clauses and promises of universal basic income are promises of this kind, made by the side whose power is rising. Powell’s analysis says the declining side does better bargaining over capabilities than over promises: equity, governance seats, control of endpoints, what the other side can do rather than what it may do.
The US–China rivalry is more than a parallel. It is one reason humanity cannot act as a single declining power, because each country’s labs are the other country’s reason not to slow down. The uncomfortable reading of the Thucydides frame is that the classical move for a declining power is preventive action while it still has leverage. Translated out of war, that means regulation and ownership claims now, rather than redistribution promises later.
2 Further reading
- Kulveit et al. (2025) and the gradual disempowerment site
- My notebook: gradual disempowerment, human domestication, economics of cognitive and labour automation, spamularity and slop
- So you’ve joined a union, on what I learned as a workplace delegate
2.1 Community notes
TODO
3 Interesting orgs
-
IT for Change aims for a society in which digital technologies contribute to human rights, social justice and equity. Most mainstream approaches to digital technologies for development are either techno-utopic or promote a market-fundamentalist approach, often both together. A digital makeover of economies and societies has become a convenient opportunity for wholesale corporatisation of social systems, displacing the cornerstone ethics of democratic participation, commons, social justice and gender equality from development discourse.
Fairwork Certification | Independent certification for the labour behind AI. Seems to be spun off (and specialized from) Fairwork, the original project by the Oxford Internet Institute & WZB Berlin.
-
AIAAIC (standing for “AI, Algorithmic and Automation Incidents and Controversies”) is an independent, politically and technologically non-partisan grassroots public interest initiative that examines and makes the case for real AI, algorithmic and automation transparency, openness and accountability.”
