I. After the Diagnosis
The previous paper closed a diagnostic chain: when compute, data, and model weights enter the functional transformation interval, they slide from tools into control variables; this happens on each of the three carriers and they reinforce one another; the result is a three-class structure; hyper-concentration distorts rules through institutional capture, constituting pro tanto injustice; and when distortion reaches cognitive exclusion, the injustice becomes a violation of fundamental rights. The diagnosis stopped at diagnosis. That was the honest thing to do.
But the question after diagnosis cannot be evaded. What are the dispossessed to do?
If the hyper-concentration of the means of cognitive production has lost moral protection, is it legitimate to reclaim from it the conditions for participating in cognitive production? If legitimate, how do we keep that reclamation from sliding into disorganized technological Luddism? If illegitimate, or legitimate but impracticable, where is the way out?
These questions push the argument out of the theoretical space of the previous paper. The moment action is touched, the tools used there—industrial organization, public choice, Rawlsian distributive justice—begin to falter. They can diagnose the injustice of institutions. Under the premise that institutions have been deeply captured, their capacity to guide how the dispossessed should act rapidly diminishes.
What takes over is the older tradition of collective action and class analysis: not because it is more "correct," but because it is the tradition that has most systematically developed the proposition of how the oppressed change structures through collective action. There is an additional difficulty here that did not appear in the comparable paper of the Wealth Trilogy: the dispossessed in the AI era are harder to organize than traditional workers. That difficulty is one of the central problems of this paper.
II. The Legitimacy of Technological Exit, and Its Dead End
First grant the premise: the dispossessed have the right to refuse participation in an unjust structure.
This right follows directly from the diagnosis. If hyper-concentration already constitutes pro tanto injustice—algorithmic rents seizing co-production surplus, institutional capture distorting rules, cognitive exclusion depriving people of participatory rights—then refusing this structure is not breach of contract or free-riding. Refusal of a structure that has lost moral protection does not constitute moral fault, any more than resistance to a regime that has lost legitimacy constitutes rebellion. In the AI context the refusal takes many forms: refusing a closed-source model, refusing to contribute data, refusing to build on a particular base, turning to open source, withdrawing from AI use altogether.
The problem is not with the logic. The problem is with practice.
2.1 The Triple Dilemma of Individual Exit
Individual technological exit—one person or a small group refusing to use a particular AI service—faces a triple dilemma.
Asymmetric costs. Free inference is the core tool of incorporation: it eliminates the price barrier to use, making refusal irrational. A cognitive serf who refuses mainstream AI loses the cognitive gains with which peers compete. In a market where AI capability increasingly becomes a prerequisite for employment, education, and information access, the exit-er does not preserve purity; the exit-er is eliminated by competition. Compute lords lose nothing from individual exit; the exit-er loses a living space. Each person, after rational calculation, chooses to keep using—even knowing that use structurally incorporates them.
Exit does not touch the structure. Even if some individual bore the cost, their exit changes nothing. Compute remains concentrated, the data flywheel keeps spinning, weights stay sealed, the three-class structure stands. Individual exit is a gesture expressed within an unjust structure; the structure does not budge. It redistributes a few positions within an unjust structure. It is a symptom, not a cure.
Exit intensifies marginalization. The cost of exit falls most heavily on the most vulnerable. A cognitive serf with abundant resources—highly skilled, with savings, with alternatives—might bear exiting a closed-source ecosystem. A barely-sustained serf, or a cognitively excluded person, cannot afford any form of exit. They struggle even to participate, let alone to exit. Individual exit as a "strategy of resistance" is open only to a small privileged group; far from mobilizing the dispossessed, it stratifies them. Those who can exit do; those who cannot are forced to stay; the gap widens.
2.2 The Modern Luddism
If individual technological exit ever took collective form—a "refuse AI" movement—it would degenerate into a modern variant of Luddism.
The nineteenth-century Luddites smashed looms because they correctly identified machines as instruments that deprived them of livelihood. But smashing machines does not change capitalist relations of production: capitalists buy new machines, while Luddite workers are hanged or exiled. Their tragedy was not that their anger lacked justification. It was that they directed the attack at technology rather than structure. They smashed the symptom, not the cause.
"Machine-smashing" in the AI era takes new forms: refusing to use AI, demonizing AI, demanding bans, attacking infrastructure. The anger comes from real dispossession, so it has emotional legitimacy; but it is structurally ineffective. Compute lords will not abandon concentration because some people refuse AI, any more than nineteenth-century capitalists abandoned the factory system because looms were smashed. Worse, a "refuse AI" movement can be exploited: it lets opponents be portrayed as anti-progress Luddites, consolidating the narrative that AI is inevitably dominated by a few large corporations.
The direction is not to reject technology but to change the property rights structure of technology. The nature of the problem does not permit resolution at the individual level, nor at the disorganized level. It demands resolution at the structural level.
III. The Boundaries of Liberal Theories of Action
Liberalism has a rich theory of action in the AI domain: consumer choice and market competition; Schumpeterian innovation disruption (IBM disrupted by Microsoft, Microsoft by Google); regulatory reform through antitrust, data protection, AI safety legislation; Ostromian open-source community self-governance (Hugging Face, EleutherAI, LAION); open technical standards and interoperability to reduce migration cost.
Denying that these paths exist would be dishonest. The problem is the premise each depends on.
Consumer choice fails under lock-in. When data flywheels make a particular model the de facto standard, users cannot migrate because of the quality gap. Migration is not switching accounts; it is accepting a significant capability downgrade. In a market where AI capability is increasingly critical, that means accepting competitive disadvantage. The choice is formally present but substantively empty: you "can" leave, but the cost renders the choice meaningless.
Innovation competition fails under the kill zone. Players with overwhelming compute and data advantages can rapidly replicate any startup's innovation and kill it before market position is established. The historical analogies (IBM to Microsoft to Google) work only because the barriers to entry in that era were far lower: training a useful software product did not require billions of dollars in compute, whereas training a frontier AI model does. The magnitude difference in barriers greatly diminishes the force of the analogy.
Regulatory reform fails under institutional capture. Compute lords, through lobbying, revolving doors, and setting high compliance thresholds in the name of "safety," have captured AI regulatory legislation. A captured regulator cannot effectively constrain its captors. Regulatory reform can improve local conditions; under deep capture it is insufficient to change the structure that produces the capture.
Open-source self-governance fails under the scale asymmetry. Ostrom's self-governance model applies to local, limited-scale common-pool resources. Against global-scale AI infrastructure, open-source communities rely on voluntary labor and limited funding; compute lords command tens of billions in R&D. Open source can provide alternatives—a condition for resistance—but it cannot dismantle concentration, because the compute scale required to train frontier models exceeds the capacity of any open-source community.
Technical standards fail under ecosystem binding. Compute lords have strong incentives to resist open standards, promoting proprietary standards through ecosystem tool binding, API incompatibility, and non-transferability of fine-tuning assets. Open standards can only be established when multiple parties with sufficient power jointly push for them—and forming such a coalition is itself a collective action problem.
The shared premise-level cause is that all of these presuppose an institutional environment that has not been deeply captured: competitive markets, surmountable barriers, independent regulators, parity of scale, willing cooperation among multiple parties. The previous paper's diagnosis challenges these presuppositions. When concentration of the means of cognitive production is systemic—when the phase transitions of compute, data, and weights have all occurred, when institutional capture spans legislation, regulation, public discourse, talent, and standards—the efficacy of liberal theories of action rapidly declines. They can improve conditions; they cannot change the structure that produces the conditions. Diagnosis is their strength; action guidance under deep capture is not.
IV. The Cognitive Coordination Trap
Liberal theories of action lose efficacy under deep capture. But there is an even more fundamental question: why don't the dispossessed unite? Not because they think the system is fine, but because unity itself faces a structural dilemma.
The argument of this section builds on a real tradition. Vallas & Schor (2020) identified the general organizational difficulties of platform labor: platforms render the labor process more "legible" to employers than to workers, enabling employers to "actively discriminate against their workforce." Prassl (2018) systematically analyzed the collective-rights dilemmas of platform workers. The repression and channeling mechanisms in Tilly's political process model, and Earl's (2022) research on digital repression, both give mature frameworks for the suppression of collective action in platform contexts. The first two components below—spatial dispersion and infrastructure capture—are general features already articulated in this literature. The contribution specific to this paper is to identify the dimensions on which AI annotation labor is specifically intensified, compared to the general gig economy—employer visibility and leverage points—and to integrate those difficulties into a nameable, testable structure.
Call this structure the Cognitive Coordination Trap (CCT): four mutually constraining components, the first two general to platform labor, the latter two structurally intensified in AI annotation labor. Together they form a structure of multiple binding constraints—when multiple constraints bind at once, relaxing any single one does not improve the outcome.
4.1 The Classic Paradox, and a Broken Premise
Olson revealed a counterintuitive truth: when a group's interests are harmed, each individual has an incentive to wait for others to take the lead while free-riding on the result. Granovetter's threshold model deepens the dilemma: each person has a threshold for joining, and if even the person with the lowest threshold will not go first, the chain never starts.
These are classic dilemmas of collective action. They apply to all dispossessed groups. But classic theory assumes one premise: the dispossessed can be organized, they just choose not to organize. The solution of the threshold model is "someone must go first." Once someone breaks the silence, the chain begins. The history of traditional labor movements seems to confirm this: organizers step forward, workers follow, strikes spread, institutions respond.
In the AI era that premise is shaken. The problem is no longer that the dispossessed choose not to organize, but that the organizational conditions that would allow them to unite are eliminated by the concentration structure of the means of cognitive production. This is not an intensified version of the classic dilemma. It is a structurally different dilemma: a trap.
4.2 The Four Components
Component I: Spatial dispersion. Traditional labor movements were possible because workers were physically concentrated in factories—same space every day, knew each other, shared labor experiences. Capital's concentration of workers in production sites ironically created conditions for workers' collective action. Cognitive serfs lack this condition. Their workplaces are dispersed: each uses AI services before their own computer, builds applications in their own company, contributes data in their own life. A developer using GPT-4 in Beijing, one in Bangalore, one in Berlin have no daily interaction, no shared space, no visible shared identity. The dispersion of the lowest stratum—data annotation workers—is extreme: each works at a network node in their own home, not even knowing who else is annotating similar tasks for the same platform.
Component II: Infrastructure capture. Collective action requires infrastructure. Traditional labor movements relied on union halls, workers' newspapers, assembly spaces—politically suppressed, perhaps, but not physically controlled by employers. In the AI era, the primary infrastructure that cognitive serfs can use for organization—digital platforms, communication tools, social networks—is in the hands of the targets of collective action or their allies. You cannot use the platforms of compute lords to organize actions against compute lords. Even where platforms do not actively censor (in most cases they do not, because resistance is too small to warrant it), their visibility control over organizing actions—algorithmic recommendation, content moderation—is itself a latent constraint. Annotation workers coordinate in WhatsApp groups; those groups run on the cloud services of compute lords, using infrastructure funded by compute lords.
Component III: Incorporation without trigger points. Traditional labor movements were often triggered by visible oppressive events: layoffs, pay cuts, workplace injuries. These events provided emotional fuel and moral legitimacy: "they cut our wages by 20%" is a powerful rallying cry because it points to a concrete, visible, attributable injustice. The "dispossession" of cognitive serfs lacks such trigger points. Free inference is not oppression; it is a service. Data contribution is not expropriation; it is usage. Ecosystem lock-in is not imprisonment; it is convenience. Every service provided by compute lords improves the immediate condition of cognitive serfs even as it consolidates structural dependence. Annotation workers do have visible trauma—psychological injury, extreme low wages—but the trauma is diluted across the four layers of the outsourcing chain (worker—contractor—platform—end client), so no one claims responsibility; no single "concrete employer" can be pointed at.
Component IV: Identity opacity. Collective action requires a shared identity: "we are workers." But "cognitive serf" is not yet a recognized identity. Most AI users see themselves as consumers or users, not as serfs. The collective "we" cannot be constructed, and so there is no subject for collective action. The illusion of mobility dissolves identity further: if everyone believes that "as long as I work hard I can become a lord," no one identifies as a serf.
4.3 The Multiple Binding Constraints Structure
The crux is that the four components bind simultaneously, and in economics "multiple binding constraints" names exactly this case: when several constraints bind at once, relaxing any single one does not improve the outcome; only relaxing sufficiently many at once will change the system's state.
The relationship among the four is not causal feedback ("break one and another re-closes") but simultaneous binding:
- Spatial dispersion constrains identity construction (without meeting, a "we" cannot form)
- Identity opacity constrains interest recognition (without identifying as serfs, shared interests cannot be seen)
- Absence of trigger points constrains mobilization (without visible unjust events, there is no occasion for action)
- Infrastructure capture constrains organization (even if the first three were overcome, the tools of organization are in the target's hands)
An online community connects dispersed annotation workers, relaxing spatial dispersion; but they do not see themselves as exploited workers, so the identity opacity constraint still binds, and collective action still cannot be launched. The trap's power does not come from any single component's self-repair, but from the joint constraint formed by multiple constraints binding at once. This is more precise than "self-repair" and more amenable to empirical testing: if the trap is a structure of multiple binding constraints, then effective resistance must exert pressure on several constraints at once—which is the design logic behind the two levels of resistance strategies in Section V.
4.4 Empirical Corroboration
The trap is not pure theoretical deduction. All four components have observable empirical correlates.
Spatial dispersion: In May 2023, approximately 150 African workers supporting systems such as Facebook, TikTok, and ChatGPT gathered in Nairobi and pledged to establish the African Content Moderators Union (ACMU)—the most significant organizational attempt in AI annotation to date. 150 people, against millions of annotation workers globally, is an extremely small proportion. For comparison: the unionization rate in the U.S. "Professional and Technical Services" industry (covering most tech employment) is only 1.3%, surpassed only by finance. Even for formally employed U.S. tech workers the organizational rate is already extremely low; for cross-border outsourced annotation workers the difficulty is an order of magnitude higher.
Incorporation without trigger points: The largest collective action attempt in the AI field is the "Pause Giant AI Experiments" open letter initiated by the Future of Life Institute in March 2023, which received over 30,000 signatures (including Bengio and Musk). The actual effect was zero: no major lab paused training. Expressive collective action—signatures, appeals—lacks the leverage of traditional strikes; without halting production, no pressure can be applied; the moral appeal of 30,000 people could not constrain a capital-driven AI race. Most signatories were researchers rather than the dispossessed, further confirming that even relatively organized groups, without the capacity to halt production, cannot produce substantive impact.
Infrastructure capture and exit costs: When annotation workers or developers try to migrate from one platform to another, costs are not simply "switching accounts." Industry data shows that the average cost for an enterprise to migrate from one AI base to another is approximately $315,000, with switching costs accounting for 19–34% of total AI investment, and the sunk loss of fine-tuning assets reaching $100,000 to $500,000. OpenAI's API deprecation notice in November 2025 gave developers only about 6 months to migrate. Each person is locked into their own platform assets; the cost structure of collective switching makes coordinated action unformable.
4.5 The Gig Economy as Control: What Is Specific to AI Annotation
The gig economy already shows the organizational difficulties of dispersed labor—difficulties that overlap heavily with the first two components of the CCT. California's Proposition 22 dispute is the key case: Uber, Lyft, and DoorDash spent over $224 million to pass Prop 22 (defining gig workers as independent contractors, excluding them from collective bargaining rights)—the most expensive ballot measure in California history—and in July 2024 the California Supreme Court unanimously upheld it.
But the gig economy also shows the possibility of breakthroughs. And this is precisely the contrast the CCT must explain.
In May 2026, approximately 70,000 Uber/Lyft drivers in Massachusetts formally established the first gig driver union with collective bargaining rights in the United States, covering over 800,000 drivers. If the first two components of the CCT (spatial dispersion + infrastructure capture) were sufficient to lock down collective action, Uber drivers should not have achieved this. They broke through because gig drivers are more favorably positioned than AI annotation workers on two dimensions related to the third and fourth CCT components.
First, employer visibility. Uber drivers face a single visible platform; they know "who is exploiting me" and can point to a concrete negotiating counterparty. AI annotation workers face multiple layers of outsourcing chains: worker→contractor (e.g., Sama)→platform (e.g., Scale AI/Remotasks)→end client (e.g., OpenAI/Meta); no single "employer" can be pointed at. The core difficulty of Daniel Motaung's 2022 lawsuit against Meta and Sama was precisely the ambiguity of the defendant's identity: who is responsible for the labor conditions of annotation workers?
Second, leverage points. Uber drivers have regional leverage: if drivers in one city stop accepting rides, the city's ride-hailing service collapses, exerting visible pressure. AI annotation tasks can be globally reallocated: if Kenyan workers strike, the platform transfers tasks to the Philippines or Bangladesh without interrupting production. The 2024 case of Remotasks' wholesale withdrawal from Kenya, Nigeria, and Pakistan is even starker: the platform did not passively respond to worker strikes but proactively withdrew from the entire region, instantly nullifying the organizational networks workers had built, and depriving workers of even the leverage to demand unpaid wages.
These two dimensions—the obscuring of employer visibility and the loss of leverage points—constitute the specific constraints that distinguish AI annotation labor from the general gig economy. They are not general features of platform labor (gig drivers are clearly better positioned on both). They are constraints directly tied to the specific structure of the means of cognitive production: globally allocatable annotation tasks and multi-layered intermediary outsourcing chains. The specificity of the CCT lies here—not in claiming to have discovered all mechanisms of organizational difficulty in platform labor (Vallas & Schor, Prassl, and Cant have done that systematically), but in identifying AI annotation labor's specific intensification on the two dimensions above and integrating that into a multiple-binding-constraints framework.
4.6 The Falsifiability of the Trap
The value of a theoretical mechanism depends in part on whether it can be falsified. The key testable claim of the CCT is not a definitional proposition ("the AI field is more dispersed than factories," which is near-tautological) but a comparative empirical prediction.
Prediction: Controlling for per capita income, labor market size, and labor law frameworks, the unionization rate and collective action success rate of workers in AI annotation outsourcing hubs (such as Nairobi, Manila) should be lower than that of gig drivers (ride-hailing, delivery) in the same region, with the difference primarily attributable to two dimensions: employer visibility (outsourcing chain obscurity vs. single visible platform) and leverage points (globally allocatable vs. regional work stoppages).
This prediction is falsifiable. If research finds no significant difference between the unionization rates of AI annotation workers and gig drivers, then the CCT's "AI-specificity" claim is falsified, and the difficulty would be attributed to the general structure of platform labor (spatial dispersion + infrastructure capture) rather than the specific structure of the means of cognitive production. Currently no systematic three-way comparative study exists using unionization rate as the dependent variable (AI annotation workers vs. gig drivers vs. traditional service workers). Existing data—ACMU with only 150 members, BLS tech industry unionization rate of 1.3%—provides directional support, but systematic comparative research remains to be conducted. One purpose of formalizing the CCT is to provide a testable framework for exactly that research.
4.7 The Organizational Impossibility of the Cognitively Excluded
If organizing cognitive serfs is already difficult, organizing the cognitively excluded is nearly impossible: they cannot even reach the starting point of the trap.
The cognitively excluded are excluded precisely because they lack the conditions for participating in the cognitive production-consumption cycle (language, infrastructure, economic, literacy). These are exactly the conditions necessary for organizing collective action. A person who cannot effectively use the internet cannot participate in online collective action; a person whose native language is not supported by any mainstream platform cannot be included in cross-language mobilization networks; a person struggling for basic survival has no surplus energy to participate in any form of organized action.
A cruel paradox: the group that most needs collective action is precisely the group least capable of being organized into it. The conditions of their exclusion are simultaneously the conditions that make them unable to organize resistance. The structure deprives the excluded of the conditions for participation, and the deprivation of participation conditions in turn makes it impossible for the excluded to organize to change the structure.
This closed loop has no purely logical solution. It can only be partially, incrementally, and externally broken—by those not yet fully excluded (sympathizers among cognitive serfs, international NGOs, members of marginalized language communities who have already achieved some degree of digital access) fighting for visibility and conditions for the excluded. This is precisely the significance of the data contribution strategy in the first level of Section V.
4.8 The Trap as the Deepest Protective Layer
The Cognitive Coordination Trap is the deepest protective layer of the concentration structure of the means of cognitive production. The previous paper diagnosed the explicit mechanisms: institutional capture, public discourse shaping, standard definition. The Wealth Trilogy added hegemony (making people not question), exhaustion (making people have no energy), and coercion (making people not dare). Beneath all of these lies the trap: even if a dispossessed person questions the structure, has the energy to resist, and dares to resist, the coordination trap still makes them choose to wait. No one needs to design this mechanism. It operates automatically as long as the dispossessed are each in a state of dispersion, without identity, without trigger points, without autonomous infrastructure. Compute lords do not need to actively divide the dispossessed. The dispossessed are already unable to unite because of the trap's structure.
Two auxiliary narratives reinforce it. The illusion of mobility—the boundaries of the three-class structure appear fluid ("as long as you work hard you can become a lord")—dismantles the cognitive basis of class solidarity, directing blame at oneself rather than the structure. The narrative of technological inevitability—"AI will naturally be dominated by a few large corporations because training frontier models requires enormous compute"—naturalizes a social structure as a technological law, making resistance appear irrational. The previous paper refuted the factual basis of both narratives (compute concentration is not inevitable; the zero-copy-cost nature of weights means that openness has virtually zero technical cost). But the power of narratives lies not in their truth but in their credibility.
In traditional labor-capital relations, workers at least had the organizational foundation of physical concentration—the factory—and coordination dilemmas could be partially overcome by union organization. In the AI context, cognitive serfs lack even that foundation: the four components of the trap jointly eliminate the very possibility of organization. This is the structural reason why the dispossessed in the AI era find it harder to resist than traditional workers.
V. Two Levels of Resistance
If the trap involves multiple binding constraints, then effective resistance cannot be a single strategy. It must be multi-level. The strategies below fall into two levels, differing in nature: the first plays the game within the rules defined by compute lords; the second creates a production entity of a different nature. The first has a ceiling. The second tries to break it.
5.1 Level One: Within-Structure Strategies
The three strategies that follow share a feature: they all operate within the framework of the means of cognitive production controlled by compute lords, using cracks in the structure to secure limited space for the dispossessed. They are not emancipatory plans. They are pragmatic strategies to improve conditions without changing the property rights structure.
(I) Data contribution: breaking the invisibility of the cognitively excluded. Target: the cognitively excluded, especially low-resource language communities. Their primary dilemma is not "using poorly" but "not being seen at all"—their languages, cultures, and needs barely exist in training corpora. By participating in projects such as Common Voice (Mozilla's crowdsourced speech dataset), Masakhane (African language NLP), and OPUS (open parallel corpora), marginalized communities can bring their languages into public training corpora. These contributions are slow and intergenerational—a language may require years of collective annotation work to move from "invisible" to "having basic representation." But it is the prerequisite for structural loosening: without presence in the data, any subsequent fine-tuning, adaptation, or service is impossible.
Boundary: this strategy depends on community organizational capacity and collective will, and does not solve immediate service gaps. There is also a risk—if data contributions only enter public corpora, they may be scraped for free by compute lords to train closed-source models, without giving back to contributing communities. Data contribution therefore needs to be combined with community data governance (data cooperatives, usage condition agreements). But even without perfect governance, existential contribution has value: it at least breaks total invisibility.
(II) Community fine-tuning: contesting adaptation on open-source bases. Target: communities whose languages or domains have some representation, however imperfect, in open-source base models. Through lightweight fine-tuning techniques such as LoRA, communities can adapt open-source bases to their own languages and domains. An open-source model that performs well in English but poorly in Swahili can be fine-tuned with community-organized annotation data to improve performance on specific tasks.
The limits of this strategy as resistance must be honestly acknowledged. Not all fine-tuning on open-source bases constitutes resistance. Most fine-tuning on HuggingFace is developers building applications on Llama and publishing to commercial markets—which serves more to strengthen Llama's ecosystem stickiness for Meta than to challenge the structure of compute lords. The criterion is: resistant fine-tuning fills service gaps that compute lords are unwilling to provide (low-resource languages, de-censored dialogue, cultural adaptation for specific communities); dependent fine-tuning increases the commercial ecosystem value of open-source bases, whose greatest beneficiaries are the large companies that trained the base.
Regardless of type, all such fine-tuning faces a deep constraint: as long as workers lack the capacity to train models from scratch, they must rely on a large company that has that capacity to provide the base, and hope that this company maintains an open-source strategy. Community fine-tuning—even resistant—is always a dependent activity, its preconditions set by the competitive strategies of big companies, not by community control. Big companies can change open-source strategy at any time (HashiCorp, Redis, and MongoDB have already shifted from open source to "source available"), at which point all fine-tuning assets based on old versions face devaluation. Empirical research also shows that the effectiveness of community fine-tuning depends on the quality of the base model's pre-training representations in the target language. For nearly invisible languages, fine-tuning only produces "more confident noise": LoRA cannot create non-existent representations out of thin air. Data contribution (I) and community fine-tuning (II) are therefore complementary.
(III) Serf competition: contesting portability among lords. Target: the more technically capable among cognitive serfs—developers, enterprise users, AI engineers. Not refusing to use lords' bases (that is a dead end), but exploiting competition among lords to force them to release more fine-tuning rights and ecosystem openness.
Mechanism: lords compete for ecosystem dependence (developers, enterprise customers, data contributors). To attract that dependence, a particular lord may proactively open up more—similar to google open-sourcing Android to counter Apple's vertical integration. Serfs, acting as "power users," can build portable fine-tuning toolchains (llama.cpp, Ollama, vLLM, and other unified inference frameworks) to make their fine-tuning assets transferable across bases, forcing more lords to follow suit with openness. The low-price open-source strategies of Chinese players such as DeepSeek, Alibaba's Tongyi Qianwen, and Zhipu have already exerted objective price and openness pressure on OpenAI and Anthropic—real-world corroboration of the logic of serf competition.
Boundary: this strategy depends on a key premise, that lords have not reached a tacit agreement of total closure. If major compute lords form a jointly closed tacit agreement—jointly restricting fine-tuning, jointly raising migration costs, jointly resisting open standards—the foundation of this strategy collapses. Currently commercial competition tends toward some degree of openness, but this is a dynamic, reversible condition.
The shared ceiling. All three strategies share a structural limitation: they operate within the property rights framework defined by compute lords. Data contributors' data may be used for free by lords; community fine-tuners depend on bases released by lords; serf competitors depend on the competitive landscape among lords. No matter how clever the strategy, efficacy ultimately depends on the strategic choices of compute lords, not on the dispossessed's own control. This is the insurmountable ceiling of within-structure strategies: without changing the property rights structure of the means of cognitive production, resistance can only contest space, not change the rules.
5.2 Level Two: Alternative Production Entities
The shared ceiling of Level One points toward a more fundamental direction: not playing the game on the bases of compute lords, but creating a production entity of a different nature—an AI infrastructure not driven by profit maximization, owned by the public or community, with returns flowing back to co-producers. This is not contesting space within the structure. It is an attempt to change the structure itself.
As long as the property rights structure of the means of cognitive production remains unchanged, the strategic efficacy of the dispossessed is always constrained by the choices of compute lords. Data contribution, community fine-tuning, and serf competition are all valuable, but their ceiling is structural. The only path to break it: beyond the compute lords, establish a new cognitive production entity whose property rights structure prevents algorithmic rents from being monopolized and channels them back to co-producers. This is precisely the direction in which the ρARA (algorithmic rent allocation rate) proposed in the first paper needs to be fundamentally raised at the institutional level.
A genuinely resistant alternative production entity must satisfy three normative conditions:
- Ownership attribution. The entity is held by the public or community, not by private shareholders. This is different from "low price" or "open source": low price is a pricing parameter (any competitor can temporarily adopt it), open source is a licensing strategy (any large company can choose or abandon it), while ownership attribution is a fundamental arrangement of the property rights structure. An entity's "publicness" does not lie in whether it is low-priced (DeepSeek's low prices are a competitive strategy), but in who holds its residual claim—whether algorithmic rents flow to all owners (the public/community) or to a small number of shareholders. This is the property rights foundation of ρARA.
- Democratic governance. The entity is governed not by capital dictatorship (shareholders' meeting to board to management) but by the participation of diverse stakeholders: owners (the public), technical experts, workers, user representatives. The difficulty is the tension between democratic participation and decision-making efficiency. The pace of AI frontier competition is extremely fast, and multi-stakeholder democratic processes may be slower than private competitors in decision speed. This tension cannot be resolved by abandoning democratic governance (which falls back to traditional "administrative centralization" public ownership); it must be mitigated through honest positioning of technical goals.
- Return flow. The entity's operational returns (after reinvestment) do not flow to shareholders in the form of profits but flow back to co-producers—the public, data contributors, annotation workers—in the form of universal dividends or public services. This elevates ρARA from near zero to a level significantly higher than that of private AI enterprises.
Honest positioning of technical goals. Alternative production entities do not need to win the AI arms race. An honest goal is to catch up to the frontier rather than lead it: to reduce training and service costs through efficiency innovation rather than scale accumulation, to maintain "adequate and not lagging" capability in key areas, to prioritize serving areas neglected by closed-source giants (low-resource languages, public service sectors, marginalized communities). DeepSeek has already shown that through architectural innovation (MoE, multi-token prediction), "near-frontier" capability can be achieved at costs far below those of frontier model training. The goal is not to replicate OpenAI. It is to prove that under a publicly owned property rights framework, cognitive production need not obey the logic of capital valorization.
Historical precedents. Public broadcasting (BBC, NHK) provides public services beyond commercial media, changing the competitive nature of the media ecosystem. Public utilities (the universal service principle for electricity, water, telecommunications) prevent private extraction in natural monopoly sectors through regulated or publicly owned forms. In sectors with strong economies of scale and public goods characteristics, the existence of public entities changes the competitive nature of the entire market—not because public entities are always more efficient, but because their existence prevents private entities from extracting without restraint. AI infrastructure has similar economies of scale and public goods characteristics; the existence of alternative public AI entities can play the same counterbalancing function.
Relationship with Level One. Level Two does not replace Level One; it gives Level One a structural anchor. Data contributors' labor has a genuine home in alternative entities (data flows back to communities rather than being scraped for free by lords); community fine-tuners work on publicly owned bases (bases are publicly owned and will not suddenly change open-source strategy); serf competitors gain stronger negotiating leverage (the existence of alternative entities means that "exit" no longer means just another lord, but an entity of a different nature). Level One strategies gain a higher ceiling with the existence of Level Two.
Conditions and boundaries. The establishment of alternative production entities faces difficulties far greater than any Level One strategy. It requires political power to advance property rights restructuring (cannot be accomplished spontaneously by individuals or small communities); it must address the risk of institutional capture (managers of public entities may be lobbied or bought by compute lords; the anti-capture paradox of the third paper applies here); and it must find a viable balance between democratic governance and technical efficiency. These difficulties are real. The significance of alternative production entities lies not in their ease of realization but in their pointing toward a fundamentally different direction: not playing within the lords' rules, but changing the rules themselves. The excessively low ρARA diagnosed in the first paper and the three-principle framework proposed in the third paper both require some institutional carrier for realization—alternative production entities are the embryonic form of such a carrier.
VI. Solidarity
The two levels of resistance differ in nature, but they share a premise: both require collective action. Data contribution requires community collaboration; community fine-tuning requires collective annotation; serf competition requires the widespread adoption of standardized toolchains; alternative production entities require political power to advance property rights restructuring. Not a single resistance strategy can be realized through purely individual action.
6.1 The Logical Necessity of Solidarity
Individuals cannot change structures. But individuals are not isolated: those dispossessed within the concentration structure of the means of cognitive production are not one person but millions. Dispersed, they have no power; united, they constitute a force within the structure capable of countering institutional capture by compute lords.
Solidarity is therefore not a moral appeal but a structural imperative: in a system that structurally deprives individuals of the capacity for resistance, the only possible lever is collective action. One person cannot refuse unjust rules; one million can. One person cannot make institutions respond; a social movement can. In the AI context the objects of solidarity expand to cognitive serfs and the cognitively excluded. The form of their dispossession differs—not wages being suppressed, but the means of cognitive production being monopolized—but the logical structure of solidarity is the same: dispersed individuals are powerless to change structures; a united collective is the only force capable of countering the self-reinforcing cycle of institutional capture.
6.2 The Solidarity Dimension of the Trap
The Cognitive Coordination Trap is a correction to any optimistic reading of the necessity of solidarity. Solidarity is logically necessary but practically difficult, and the difficulty is another expression of the trap. The four components are not only obstacles to collective action but obstacles to solidarity: spatial dispersion prevents the construction of trust networks, infrastructure capture places the organizational tools of solidarity in the hands of the target, incorporation without trigger points eliminates the emotional occasions for solidarity, and identity opacity makes the construction of "we" impossible.
The difficulty of solidarity in the AI era does not need to be re-argued; it is the solidarity dimension of the trap. What truly needs to be asked is: under the constraints of the trap, in what forms is solidarity possible?
6.3 Possible Forms
Despite the weight of difficulties, solidarity is not impossible. Its possible forms may be non-traditional.
Cross-stratum cognitive serf alliances. Highly skilled developers, enterprise users, and ordinary AI users all share sufficient interests (all subject to compute lords' base control, API pricing, ecosystem lock-in) that they may form some kind of alliance, contesting portability and bargaining power through standardized toolchains, open-source ecosystems, industry self-regulation. Such alliances need not take the form of traditional unions; they may appear as "open-source communities + developer associations + enterprise alliances."
International digital rights movements. Like the internationalization of environmental or human rights movements, solidarity in the AI era may take the form of cross-border NGOs, academic networks, policy advocacy coalitions pushing for institutional changes such as data rights legislation, algorithmic rent distribution, low-resource language support.
Global South and marginalized community alliances. The cognitively excluded are highly concentrated in the Global South and in marginalized language communities. Their solidarity may take the form of "Global South AI sovereignty," analogous to historical resource sovereignty movements: Global South nations jointly demand greater control over the means of cognitive production (especially public data and public compute). The BRICS New Development Bank, regional AI research collaborations, multilingual model initiatives may be embryonic forms.
These forms are preliminary, fragile, and uncertain. But their existence shows that solidarity in the AI era, though difficult, is not without the possibility of germination.
VII. A Statement of Position
At this point the argument has moved from the immanent critique of mainstream frameworks in the previous paper to a place those frameworks cannot contain. Industrial organization economics, public choice theory, Rawlsian distributive justice were sharp enough to cut through whether concentration of the means of cognitive production is just. On what is to be done, they fall silent. Liberal theories of action lose efficacy under deep capture, individual exit is a dead end, and the Cognitive Coordination Trap locks down the organizational conditions for collective action.
The narrative of collective action and solidarity takes over from here—not because it is more "correct" (no narrative can claim that), but because, in the experience of history, it is the only tradition that has systematically developed the proposition of how the oppressed change structures through collective action.
This is not persuasion. Those who agree—those who have already felt the oppression of concentration—do not need an argument to tell them solidarity is necessary; their experience of use has already taught them. Those who do not agree—those who benefit from the existing structure or believe it can self-correct—will not be persuaded, because they do not believe the structure needs to change.
What is said here is a statement of position. If you have followed the logic of this paper to this point—acknowledging the injustice of the phase transition of the means of cognitive production, the systemic nature of institutional capture, the dead end of individual exit, the severity of the coordination dilemma—then you should not pretend that all of this has no implications for action by the dispossessed. The two levels of resistance point to starting points of action for groups in different positions. Institutional change requires solidarity; solidarity requires organization; organization requires action. This paper cannot complete that work. It can at least honestly point toward it.
VIII. Conclusion: The Conditionality of Action
In the AI era, individual resistance is a dead end, collective action faces the Cognitive Coordination Trap—a structure of multiple binding constraints—and the direction for breaking through the trap's ceiling is the creation of alternative production entities.
The argument supporting this: individual technological exit does not touch the structure (Section II); liberal theories of action lose efficacy under deep capture (Section III); the Cognitive Coordination Trap structurally eliminates the organizational conditions for collective action (Section IV). The originality of the trap does not lie in discovering the general organizational difficulties of platform labor (addressed by Vallas & Schor, Prassl, Cant) but in identifying the specifically intensified constraints of AI annotation labor on two dimensions: employer visibility and leverage points. The contrast between the 70,000 Massachusetts Uber/Lyft drivers establishing a union in 2026 and the largest organizational attempt by AI annotation workers to date involving only 150 people provides preliminary support for the trap's falsifiable prediction.
Resistance strategies therefore organize into two levels. Within-structure strategies—data contribution, community fine-tuning, serf competition—face an insurmountable ceiling: as long as the property rights structure remains unchanged, the efficacy of resistance is always constrained by the choices of compute lords. Alternative production entities point toward changing the property rights structure itself. The two levels are complementary, and their common premise is collective action and solidarity.
This paper is the AI-era counterpart of the second paper of the Wealth Trilogy, The Legitimacy of Reclamation and the Necessity of Solidarity. It advances on two layers. Relative to the Wealth Trilogy, it formalizes the special difficulties of collective action in the AI era into a nameable structure of multiple binding constraints—the Cognitive Coordination Trap. The Wealth Trilogy identified coordination dilemmas (Olson, Granovetter) but treated them as general dilemmas applicable to all dispossessed groups; this paper argues that, in the AI context, the dilemma acquires a specific structural form. Relative to platform labor research, the CCT does not claim to have discovered all mechanisms of organizational difficulty in platform labor; spatial dispersion and infrastructure capture have been systematically addressed in Vallas & Schor (2020), Prassl (2018), and Cant (2019). The CCT identifies the specific constraints of AI annotation labor on the two dimensions above and integrates them into a multiple-binding-constraints framework, generating a falsifiable comparative prediction.
This paper also faces an additional paradox absent from the Wealth Trilogy: the organizational impossibility of the cognitively excluded. The group that most needs collective action is precisely the group least capable of being organized into it. This paradox has no purely logical solution; it can only be partially broken from the outside. That gives special importance to the data contribution strategy in Level One, to external solidarity actors (international NGOs, members of marginalized communities who have already achieved digital access), and to Level Two's alternative production entities (which can provide the excluded with public AI services that do not require being "organized first" to benefit from).
The honesty of action requires two acknowledgments. First: Level One strategies are pragmatic strategies for contesting limited space, not solutions. Their ceiling is structural; as long as the property rights structure remains unchanged, efficacy is constrained by compute lords' choices. Level Two points toward a more fundamental direction, but its realization requires political power to advance property rights restructuring, and its feasibility faces severe challenges.
Second: inaction deterministically preserves the status quo. Concentration will not dissolve because it has been diagnosed; the three-class structure will not loosen because it has been identified; ρARA will not rise because it has been proposed; the Cognitive Coordination Trap will not unlock because it has been named. All change requires the practice of actors—starting from the strategy in Level One closest to one's own position, while accumulating conditions and organizational experience for the institutional changes of Level Two. Begin by recognizing one's own structural position, contest space within the structure, while accumulating political conditions and organizational capacity for the alternative production entities that transcend it.
The path out of the conditional dilemma must be walked by the actors themselves. As for how to design new institutions that will not be captured after walking out—that is the question for the third paper.