The Specific Event

Meta's elite AI research unit recently attempted to harvest employee data to train its models. The effort collapsed after an internal leak triggered a worker revolt that forced the initiative to a halt. This is not a story about data privacy in the abstract. It is a story about what happens when an organization treats its own workforce as an algorithmically extractable resource, and what the structural conditions are that make that framing both tempting and ultimately self-defeating.

The Inversion Problem in Internal Platform Deployment

Classical organizational theory assumes that authority flows downward and that employees accept the terms of their participation in exchange for compensation. What Meta's situation reveals is that this assumption breaks down when the thing being extracted is not labor in the conventional sense but epistemic content - the behavioral traces, communication patterns, and cognitive outputs of knowledge workers. The employees who revolted were not objecting to being managed. They were objecting to being mined. That distinction matters considerably for organizational theory.

The ALC framework I work with in my dissertation research identifies a coordination inversion that occurs on algorithmic platforms: rather than deploying pre-existing competence into a structured environment, workers develop competence endogenously through participation in the platform itself (Kellogg, Valentine, and Christin, 2020). Meta's failed initiative represents a further inversion. The organization attempted to extract the latent competence of its workers not to improve coordination for those workers, but to build systems that could eventually replace or reduce dependence on them. The workers recognized this and acted accordingly.

Folk Theories and Structural Schemas Inside the Firm

There is something theoretically interesting about the fact that a leak triggered the revolt. This suggests that employees did not have accurate structural knowledge of what was being built or why. Their response was reactive rather than anticipatory. This maps onto the distinction between folk theories and structural schemas that is central to algorithmic literacy research. Folk theories are impressionistic and individual; they emerge from surface-level cues and are often inaccurate (Gagarin, Naab, and Grub, 2024). A structural schema, by contrast, represents genuine understanding of how a system is organized and what its constraints produce.

The employees at Meta's AI unit were operating on folk theories about how their data might be used internally. The leak converted a folk theory situation into a structural schema situation, and behavior changed immediately and collectively. This is the awareness-capability gap in a new organizational register. The gap here was not between knowing algorithms exist and knowing how to respond to them. It was between knowing that data collection was occurring and understanding the architectural purpose that collection served within Meta's model development pipeline.

What Organizational Theory Predicts About This Failure

Rahman's (2021) concept of the invisible cage describes how platform firms exert control through informational asymmetries rather than direct supervision. The cage works precisely because workers cannot see its full structure. Meta's situation suggests that this dynamic does not disappear when the platform firm turns its gaze inward toward its own employees. If anything, it intensifies, because knowledge workers have both the analytical capacity to recognize asymmetric information arrangements and the organizational leverage to resist them when those arrangements become visible.

Sundar's (2020) framework on machine agency is also relevant here. As AI systems acquire greater perceived agency within organizations, the question of whose interests those systems serve becomes harder to obscure. Meta's employees likely understood, at least implicitly, that a model trained on their behavioral data would encode their expertise in a form that the organization could deploy without their ongoing participation. The revolt was not irrational. It was a structurally coherent response to a clearly identified threat to occupational irreplaceability.

The Broader Implication for AI Governance Inside Organizations

What this episode demonstrates is that internal AI development faces a legitimacy constraint that external product development does not. When Meta trains models on public data, the affected parties are dispersed and largely unorganized. When it attempts to train models on employee data, the affected parties are present, informed, and capable of collective action. Organizations developing AI internally will need to develop governance frameworks that account for this asymmetry, not because workers will always win these disputes, but because the cost of losing worker trust in knowledge-intensive environments is substantially higher than the cost of acquiring training data through other means.

The failure here was not technical. It was a coordination failure rooted in an organizational schema mismatch. Meta's leadership appear to have been operating with a topographic understanding of internal data - knowing what data existed and where - without a topological understanding of the constraints that worker identity and occupational interest place on how that data can be legitimately mobilized. Knowing the shape of a resource and knowing the constraints governing its use are not the same thing. This distinction, which is central to my own framework, turns out to have significant practical consequences.

References

Gagarin, D., Naab, T. K., and Grub, J. (2024). Algorithmic media use and algorithm literacy. New Media and Society.

Kellogg, K. C., Valentine, M. A., and Christin, A. (2020). Algorithms at work: The new contested terrain of control. Academy of Management Annals, 14(1), 366-410.

Rahman, H. A. (2021). The invisible cage: Workers' reactivity to soft control in platform work. Administrative Science Quarterly, 66(4), 1003-1043.

Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of human-AI interaction. Journal of Computer-Mediated Communication, 25(1), 74-88.