When the Product Is the Worker's Body

This week, an AI training startup called Shift announced it would clean New Yorkers' apartments for free. The offer is straightforward in its structure: Shift sends workers into homes, those workers perform domestic labor, and the entire process is filmed to generate training data for embodied AI systems. Shift has plans to expand into London and other cities. The exchange is labor for footage, and the people whose homes get cleaned are not the ones being studied. The workers are.

The framing in most coverage treats this as a quirky data-acquisition story. I want to argue it is something more structurally significant: a case study in what happens when an organization needs to extract tacit knowledge from workers who do not themselves possess an explicit schema for what they know. This is the hard version of the competence transfer problem, and Shift's approach reveals exactly why procedural documentation fails where schema induction is necessary.

The Tacit Knowledge Problem in Embodied Labor

Domestic work involves what Hatano and Inagaki (1986) would classify as adaptive expertise rather than routine expertise. A cleaner navigating an unfamiliar kitchen does not apply a fixed procedure. She reads the topology of the space, identifies constraint structures, and improvises sequences that are sensitive to local conditions. This is not a checklist. It is a form of structural reasoning applied under uncertainty. The reason Shift needs video rather than worker interviews or written protocols is precisely because workers cannot fully articulate what they are doing. The knowledge lives in the doing.

This creates a specific organizational problem. Shift is attempting to induce schemas from behavioral observation because the workers themselves cannot provide first-person accounts that capture structural features. Rahman (2021) documented a related inversion in platform work, where algorithmic systems extract behavioral patterns from workers to constrain those same workers later. Shift's model is a lateral extension of that logic: the extracted patterns will eventually be used to train systems that perform, and potentially displace, the labor being filmed.

The Awareness-Capability Gap, Inverted

Most of my work on the Algorithmic Literacy Coordination framework focuses on workers who have insufficient schema representation of the platforms they work within. The awareness-capability gap I study is the gap between knowing an algorithm exists and knowing how to respond to its structural logic (Kellogg, Valentine, and Christin, 2020). Shift presents an interesting inversion of this problem. The workers being filmed have high capability but low explicit awareness of the structural features driving their own performance. They cannot narrate their schemas because those schemas were never formally induced - they were built through years of embodied practice.

What Shift is attempting, then, is computational schema extraction: using video and machine learning to reconstruct the structural representations that adaptive experts carry implicitly. Whether this works is an empirical question, but the organizational design assumption is revealing. Shift is betting that external observation can substitute for internal schema articulation. That bet may be wrong. Gentner's (1983) structure-mapping theory suggests that relational structure is what enables transfer, and relational structure is exactly what camera footage is worst at capturing. You can film someone wiping a counter. You cannot film the constraint satisfaction reasoning that determined which counter to wipe first.

What the Expansion Plan Signals

Shift's plan to expand into London and other cities is not primarily a geographic story. It is a signal about the organization's theory of data sufficiency. The implicit assumption is that more footage from more contexts will eventually converge on generalizable representations of domestic labor competence. This is the procedural training fallacy at scale: the belief that accumulating enough specific instances produces transferable knowledge. The ALC framework predicts this will be insufficient. What makes expertise transferable is not the volume of behavioral instances but the accuracy of the structural schema being trained against those instances.

There is also a governance dimension that Shift's announcement does not address. The workers being filmed are producing intellectual property in the form of embodied knowledge, and they are compensated with apartment cleaning rather than with equity, royalties, or schema attribution. Schor et al. (2020) documented how platform dependency creates precarity through asymmetric value extraction. Shift's model extends this into physical labor markets, and the asymmetry is even more extreme because the workers likely do not understand that their cognitive labor, not their physical labor, is the primary product being purchased.

The Structural Point

The Shift story matters because it clarifies something that purely digital platform cases obscure: the thing being extracted in algorithmic training systems is not behavior. It is the structural knowledge that generates behavior. When organizations conflate the two, they design collection systems that are comprehensive on behavioral instances and impoverished on structural features. The expansion into new cities will produce more data. Whether it produces better schemas is a separate question, and one that Shift's current framing gives no reason to answer confidently.