Meta's elite AI research unit recently attempted to harvest employee data to train its models. The effort did not survive internal scrutiny. A leak followed by a coordinated worker revolt forced the initiative to a halt before it could produce any usable training pipeline. The story has been covered primarily as a privacy scandal, but that framing misses the more theoretically interesting dimension: what this episode reveals about how organizations actually build algorithmic competence, and why the methods they choose tend to contradict what research says actually works.

The Data Extraction Premise and Its Structural Problem

Meta's approach rested on a specific assumption: that behavioral data generated by employees during ordinary work contains latent competence that can be extracted, formalized, and transferred into model weights. This is the same assumption underlying most enterprise AI training programs, just made unusually explicit. The employees generating the data are treated as repositories of procedural knowledge. The model is supposed to absorb what they know by observing what they do. Kellogg, Valentine, and Christin (2020) described a structurally similar dynamic in their review of algorithmic management at work, where firms systematically convert worker judgment into algorithmic rules without accounting for whether the judgment being codified is itself accurate or transferable. The extraction premise conflates behavioral output with underlying competence, and that conflation has practical consequences.

Why Workers Revolted, and What That Tells Us

The employee revolt at Meta was not simply about privacy, though privacy concerns were real. It was also a reaction to being instrumentalized without understanding. Workers could see that their data was being collected, but they could not see how it would be used, what features would be extracted, or whether the resulting model would make their own jobs easier or replace them. This is a textbook instance of what I have been calling the awareness-capability gap in my dissertation research on Algorithmic Literacy Coordination. Workers were aware that an algorithmic process was being applied to them. That awareness did not give them any meaningful capacity to respond effectively, because awareness and structural understanding are not the same thing. Gagrain, Naab, and Grub (2024) make a similar distinction between surface-level algorithm awareness and the deeper schema-level understanding that actually shapes behavior. Meta's employees had the former. The opacity of the initiative denied them the latter.

The Folk Theory Problem Inside the Firm

What makes this episode particularly interesting from an organizational theory perspective is that the dysfunction ran in both directions. The employees lacked structural understanding of what Meta intended to do with their data. But Meta's AI unit also appears to have operated on a folk theory of organizational competence, one that assumed behavioral data was a reasonable proxy for the tacit knowledge that makes expert employees effective. Gentner's (1983) structure-mapping theory is useful here. Genuine competence transfer requires mapping relational structures, not just surface features. Behavioral logs capture surface features: what someone typed, which files they accessed, how long they spent on a task. They do not capture the relational logic that explains why those behaviors were appropriate in context. Harvesting the data without that structural mapping is methodologically thin, and the worker revolt may have short-circuited a project that would have produced poor results regardless.

The Organizational Theory Dimension

Rahman (2021) argued that algorithmic systems function as invisible cages, structuring worker behavior through constraints that are real but not legible to the workers inside them. Meta's failed initiative inverted this dynamic in an instructive way. Here, the cage was turned inward: the organization attempted to use its own employees as raw material for a system those employees could not see or evaluate. The revolt was a legibility claim. Workers were asserting that they deserved to understand the structural purpose of what was being done with their output. That claim is consistent with what Schor et al. (2020) identified as a central tension in platform labor: the asymmetry between organizational visibility into worker behavior and worker visibility into organizational intent.

What This Means for How Organizations Build Algorithmic Capability

The practical implication is not that organizations should stop trying to build internal AI systems. It is that the extraction model of competence building is theoretically incoherent and organizationally fragile. Behavioral data without structural understanding produces systems that replicate procedures without capturing the adaptive judgment that makes those procedures effective in the first place. Hatano and Inagaki (1986) distinguished routine expertise, which handles familiar cases reliably, from adaptive expertise, which generalizes to novel conditions. Data extraction pipelines are optimized to harvest routine expertise. The interesting organizational problems almost always require the adaptive kind. Meta's workers understood this, even if they did not have that vocabulary. Their revolt was, among other things, a protest against being reduced to their procedures.