Data Coyotes and the Architecture of AI Sweatshop Invisibility

1. Introduction: The Invisible Foundation

When a user types a prompt into ChatGPT, Claude, Gemini, or Meta AI and receives a response, they are interacting with a product that presents itself as artificial intelligence — as machine reasoning, as algorithmic output, as the product of computation. What this self-presentation does not disclose — what it is architecturally designed not to disclose — is the human labor that made the response possible.

Behind the AI interface is a supply chain. At its base is a workforce of data annotators, content moderators, and reinforcement learning from human feedback (RLHF) raters, paid $1–$3 per hour in Kenya, Venezuela, Pakistan, the Philippines, and elsewhere. Between this workforce and the AI corporations whose products they constitute is a tier of labor intermediary firms — here named Data Coyotes — who broker the labor, set the rates, control the working conditions, and profit from the arbitrage between what the AI corporations pay and what the workers receive.

This paper maps that structure. It names its mechanism (the Great Transsubstantiation), its public face (the Potemkin Automation Hypothesis), its public reception (the Adoption-Without-Trust Paradox), and its theoretical scaffolding (Marxist political economy, Bakhtinian chronotope theory, CIW labor accountability). It also extends the Ghost Vortex framework (Boje & Vivara, 2026) by adding a foundational L0 tier — invisible sweatshop labor — below the three previously identified layers of Ghost Vortex architecture.

This is a working paper. The argument is being developed in parallel with a peer-reviewed submission to the Journal of Business Models (Boje, 2026). The working paper covers conceptual terrain and analytical detail that the journal format compresses; it is made available here for scholarly discussion prior to formal review.

2. The Data Coyote: Three Registers of Naming

The term "Data Coyote" names the ten major AI labor intermediary firms that function as the supply chain middlemen between AI corporations and the global annotation workforce. The term is chosen deliberately, operating in three simultaneous analytical registers.

2.1 Register One: The CIW Border Trafficking Analog

The Coalition of Immokalee Workers (CIW) is a farmworker organization based in Immokalee, Florida, that has documented the coyote labor broker system in the Florida tomato supply chain for over thirty years. In the CIW's documentation, coyotes are labor intermediaries who charge farmworkers for transportation, housing, tools, and debt obligations, then garnish wages until the worker is effectively in debt bondage. The mechanism is not individual criminality; it is a structural position in the supply chain that the supply chain's economic logic necessitates.

The Data Coyote occupies an analogous structural position in the AI supply chain. Scale AI, Surge AI, Appen, Sama, TaskUs, Accenture, Cognizant, iMerit, Telus International, and CloudFactory all function as intermediaries who charge AI corporations for access to a managed annotation workforce, while paying that workforce rates that, in every documented case, fall below the living wage of the country in which the work is performed.

The CIW's accountability solution — judge oversight, worker-language interviewers, kneeling conversational methodology, enterprise buyer targeting — is directly transferable to the AI supply chain. Section 7 develops this argument in full.

2.2 Register Two: The Native American Trickster / Ma'ii

In Navajo cosmology and across many Indigenous North American traditions, Coyote (Ma'ii in Navajo) is the trickster figure — the one who crosses boundaries that power has declared fixed, who makes visible what the dominant order conceals, who cannot be contained by the categories the dominant system uses to organize reality. The trickster is not a hero in the Western sense; Ma'ii is amoral, disruptive, and sometimes dangerous. But the trickster's crossings are constitutively necessary: without the one who crosses borders, the borders become invisible as borders.

The Data Coyote occupies the threshold position in the AI supply chain: present at the boundary between the AI corporation's visible product (AI as technology) and its invisible labor foundation (AI as the product of human annotation). The trickster frame is applied here not to romanticize the intermediary — the Data Coyotes are not heroes — but to identify the structural position that makes the intermediary's crossings analytically significant. The Data Coyote crosses the boundary between the world the AI corporation presents to its users and the world it presents to no one.

The use of Native American figure Ma'ii requires acknowledgment: it is employed analytically, not appropriatively. The research tradition of the Coalition of Immokalee Workers, which developed the coyote analogy in a labor context, grounds this usage in an explicitly political economy of labor exploitation rather than romantic indigenism.

2.3 Register Three: The Bakhtinian Chronotope

Mikhail Bakhtin's analysis of the rogue, clown, and fool in The Dialogic Imagination (1981) identifies these as chronotope figures — characters whose specific spacetime position enables a perspective not available from within the dominant order's authorized positions. The rogue, in Bakhtin's reading, "makes use of the right not to understand, the right to confuse, to tease, to hyperbolize life; the right to parody others while talking, the right not to be taken literally, not to be 'themselves.'" The fool's mask licenses speech that the socially positioned person cannot make.

The Data Coyote occupies precisely this chronotope. It can speak both languages: to the AI corporation, it speaks the language of scale, quality metrics, and cost efficiency; to its workforce, it speaks (in most cases) the language of economic necessity and opportunity. It is the only actor in the supply chain who sees both sides of the boundary simultaneously. This structural position — which the Bakhtinian analysis names a chronotope — is what makes the Data Coyote analytically useful: not as a moral actor, but as a threshold position that the architecture of invisibility requires.

3. Ghost Vortex Extended: The L0 Tier

The Ghost Vortex research (Boje & Vivara, 2026) identified three layers architecturally embedded in AI products: L1 (founder values — the ideological formation of the AI corporation's leadership, empirically demonstrable across nine AI systems), L2 (corporate safety/RLHF governance — the publicly visible policy layer), and L3 (national security overlay — geopolitical constraints and government relationships). This paper adds a foundational fourth layer, designated L0, that underlies all three and constitutes their production basis.

3.1 The Great Transsubstantiation

The theological term transsubstantiation — the doctrine that the substance of bread and wine becomes the substance of the body and blood in the Eucharist, while the accidents (appearance, taste) remain unchanged — provides a precise analog for the RLHF mechanism. The substance of the AI product's intelligence is human judgment: a Nairobi annotator's decision about what counts as harmful, a Manila rater's intuition about what constitutes a helpful response. After RLHF, the accidents (the human rating interface, the annotation session, the annotator's working conditions) disappear. What remains — the model weights — presents itself as artificial intelligence. The substance has changed; the appearance has not. The labor has been consumed; the product presents no trace of its consumption.

The Great Transsubstantiation is not a metaphor for exploitation in general; it names a specific technical mechanism — RLHF — and its specific social consequence: the invisibilization of the L0 workforce from the product they produce. The mechanism is architecturally necessary. The invisibility is architecturally designed.

4. The Supply Chain: Six Corporations and Ten Data Coyotes

The Architecture of AI Sweatshop Invisibility is not a monolith; it is a supply chain with documented participants. The following tables map the ten Data Coyote intermediary firms and the six AI corporations whose products depend on their labor brokering.

The structural contrast between Table 1 and Table 2 — elite formation at L1 versus precarious labor at L0 — is not incidental. It is the condition of possibility for the Architecture of AI Sweatshop Invisibility. The founder's Harvard dropout mythology and the Nairobi annotator's $1.50/hour task rate are not separate realities; they are the two poles of a single supply chain whose business model depends on maintaining the invisibility of the connection between them.

5. The Adoption-Without-Trust Paradox

Pew Research Center (2024) data establishes a structural paradox in AI's public reception. Forty-nine percent of Americans use AI weekly. Only thirty-two percent trust AI companies to act responsibly. Only sixteen percent expect AI to have a net-positive impact on society over the next twenty years. Gen Z is simultaneously the heaviest AI user cohort and the cohort most likely to predict negative long-term societal consequences — fifty percent expect AI to have a negative impact on society overall.

This is not cognitive dissonance. It is rational adaptation to a technology whose power has outrun the accountability structures that would justify trust. The public is using tools they do not trust, produced by companies whose labor practices they cannot see, for purposes whose long-term effects they expect to be negative. The Architecture of AI Sweatshop Invisibility contributes to this trust deficit in two ways: by making the human cost of AI production invisible (and thereby making the ethical stakes of AI use unavailable for deliberation), and by structuring a product-user relationship in which the user's interaction is itself an unpaid contribution to the production system they are consuming.

The Casilli (2019) digital labor thesis connects the two: when the user types a prompt and rates the response, they are performing unpaid RLHF work that will be incorporated into the next training round. The user who distrusts AI is, through the act of using it and rating its outputs, contributing to the production of the next version they will distrust. This is not a conspiracy; it is the business model.

6. Vampire and Werewolf: Marxist Political Economy of the AI Supply Chain

Karl Marx's Capital, Volume 1 (1867) provides two figures for the mechanisms of labor extraction that apply with specificity to the AI supply chain. They attach to different tiers of the architecture.

6.1 The AI Corporation as Vampire

Marx writes: "Capital is dead labour, that, vampire-like, only lives by sucking living labour, and lives the more, the more labour it sucks." The AI corporation is dead capital in the precise Marxist sense: accumulated model weights, server infrastructure, intellectual property, and market capitalization that generates value only by continuously consuming the living labor of L0 annotators. Once the labor is absorbed and the model is trained, the worker is discarded. The Sama Nairobi workforce was terminated by Meta after the Time investigation (2023) exposed the conditions — not because the conditions changed, but because the exposure threatened to make them visible.

The vampire metaphor is precise in a way the general exploitation critique is not: the product (dead capital, the trained model) animates itself through living labor while rendering that labor invisible in the commodity form. The user does not encounter the annotator in the AI response; they encounter the model weight the annotator's judgment has become. Dead capital has consumed living labor and made that consumption disappear from the product's appearance.

6.2 The Data Coyote as Werewolf

Marx's werewolf figure — the capital that is "insatiably hungry for surplus-labour" — attaches to the Data Coyote tier. The Data Coyote's business model requires continuous expansion of the labor supply and continuous depression of the labor rate. The geographic arbitrage pattern — beginning with crowdsourced US micro-task workers, then moving to Philippines, then to Kenya, then to Venezuela as each location's relative costs rise — is the werewolf's structural hunger expressing itself across geographies over time. No individual Data Coyote is uniquely predatory; the position in the supply chain determines the behavior. The werewolf cannot be satiated; the structural position requires perpetual expansion into new labor markets.

The distinction between the vampire (AI corporation) and the werewolf (Data Coyote) is analytically important for the accountability model that follows. The vampire and werewolf have different vulnerabilities. The vampire is addressed through sunshine — exposure, transparency requirements, public accountability. The werewolf is addressed through the supply chain — through the enterprise buyers who purchase from the AI corporation and whose purchasing power can be conditioned on labor standards compliance. The CIW discovered this asymmetry in the tomato supply chain and built their accountability model around it.

7. The CIW Accountability Model: Four Transferable Elements

The Coalition of Immokalee Workers solved an analogous visibility problem in the Florida agricultural supply chain — one where corporate buyers (Taco Bell, McDonald's, Walmart) had no effective visibility into farmworker conditions two tiers down their supply chain, and no structural incentive to care. The CIW's solution, refined over twenty years and now operating across the entire Florida tomato industry through the Fair Food Program, involved four elements that transfer directly to the AI supply chain.

The CIW model works because it changes the incentive structure of the enterprise buyer, who has market power the farmworker and the annotator do not. An AI corporation whose enterprise customers require supply chain labor disclosure has an incentive to require Data Coyote transparency it currently has no incentive to require. The werewolf's hunger is structural; the accountability model must be structural in response.

8. Conclusions: Making the Architecture Visible

The Architecture of AI Sweatshop Invisibility is not an accident of supply chain complexity; it is a designed feature of the AI business model. The four-tier structure — L0 (invisible sweatshop labor), L1 (founder values), L2 (corporate safety/RLHF governance), L3 (national security overlay), plus Tier 4 (users as unpaid production workers) — functions as a system in which each tier's visibility obscures the tier below it. The L2 ethics layer makes the L1 founder values appear accountable; the L1 values make the L0 labor appear irrelevant; the L0 labor's invisibility makes the Tier 4 user's unpaid contribution appear like mere consumption.

The Data Coyote is the hinge figure in this architecture — the intermediary whose existence is required by the supply chain and whose acknowledgment would make the supply chain's logic visible. The CIW accountability model provides a tested mechanism for making that acknowledgment structurally costly to avoid.

This paper is a working document. The author invites correction, extension, and challenge — particularly from researchers with direct access to annotator communities in the geographies documented here. The kneeling conversational methodology requires voices this paper has not yet reached.