Ghost Work
Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass (Houghton Mifflin, 2019) documented the vast, invisible workforce of humans who perform the labor that makes artificial intelligence appear autonomous — data annotation, content moderation, on-demand micro-tasks, the training and evaluation work that every major AI system depends on. Gray and Suri's five-year ethnographic research across four countries established that the labor was not peripheral to AI but foundational, that the workers performing it numbered in the tens of millions globally, and that the industry's systematic invisibilization of the labor produced working conditions the industry would not defend if they were visible. The book became the academic counterpart to what Janah had been arguing operationally for over a decade, and it provides the broader framework within which the Samasource trajectory can be understood as one instance of a structural pattern.
The Labor Visibility Trap — Contrarian ^ Opus
There is a parallel reading where making ghost work visible may not be the emancipatory move it appears. The entire framing assumes that once labor becomes legible to regulators and consumers, market and political pressure will improve conditions. But visibility in platform economies has historically enabled more sophisticated extraction, not less. When Amazon makes warehouse labor 'visible' through productivity dashboards, workers experience intensified surveillance and algorithmic discipline. When content moderation became a public concern, platforms responded by offshoring to jurisdictions with weaker labor protections while performing concern in Western markets. Visibility became a mechanism for bifurcating the labor supply chain into 'accountable' (expensive, regulated) and 'unaccountable' (cheap, invisible) tiers.
The deeper issue is that Gray and Suri's framework accepts AI systems as they are currently architected and asks how to make the labor powering them more dignified. But the architecture itself may be the problem. If AI development actually required transparent, well-compensated human labor at every stage, the economic model of most AI companies would collapse. The systems are designed around labor's cheapness and disposability. Making that labor visible and fairly compensated would not reform AI development — it would require abandoning the current approach entirely. The regulatory frameworks Gray and Suri advocate may simply formalize a two-tier system: visible, protected labor in some jurisdictions performing oversight work, and invisible, exploited labor continuing to do the actual training and evaluation in places regulators cannot reach. Visibility may be a way of managing the contradiction, not resolving it.
In the AI Story
The book's central empirical contribution was establishing, through sustained fieldwork with actual workers on platforms including Amazon Mechanical Turk, Microsoft's UHRS, and similar systems, that the invisible labor supply chain powering AI was larger, more globally distributed, and more systematically exploited than the industry acknowledged. The contribution has become foundational for subsequent research on platform labor, the Muldoon study, and emerging regulatory responses.
Gray and Suri coined the term 'ghost work' specifically to name the paradox of labor that is structurally essential to AI systems while being systematically rendered invisible by those systems' presentation to end users. The naming mattered: making the labor legible is a precondition for addressing its conditions, and the conditions cannot be addressed if they cannot be seen.
The book's policy recommendations align with the institutional architecture the Muldoon study later specified: combination of worker organization, civil society oversight, and regulatory intervention. The authors explicitly argue that market mechanisms alone are structurally incapable of sustaining dignity in ghost-work labor, because the economic logic of the supply chains pushes toward cost minimization in ways that exploitation is the predictable endpoint of.
The relevance to Janah's framework is that Ghost Work provides the systemic context within which Samasource operated. Janah was attempting to build an exception to the pattern Gray and Suri documented. The exception worked for a period, under specific institutional conditions. The broader pattern continued. The post-2020 Sama trajectory suggests that the exception, absent continuous institutional maintenance, eventually assimilates to the broader pattern — which is precisely the warning Gray and Suri's work makes legible.
Origin
The research began in 2014 as a Microsoft Research project and evolved into a five-year ethnographic study across the United States and India, with fieldwork involving direct observation, interviews, and participation in the platforms under study.
The book's publication in 2019 established it as the reference text for academic and policy engagement with AI labor supply chains, influencing subsequent research including the Muldoon study and regulatory frameworks in multiple jurisdictions.
Key Ideas
Labor invisibility as systemic feature. The industry's presentation of AI as autonomous is structurally dependent on the invisibility of the human labor powering it — invisibility that is not accidental but architected.
Scale of the supply chain. Tens of millions of workers globally participate in the ghost-work economy, numbers that exceed the direct employment of the AI industry many times over.
Structural exploitation tendency. The economic logic of platform labor systematically pushes toward exploitation absent countervailing institutional pressure — a pattern that matches what Muldoon later documented at Sama specifically.
Policy architecture. Worker organization, civil society oversight, and regulatory intervention are required together; no single component is sufficient.
Appears in the Orange Pill Cycle
Visibility as Necessary Precondition — Arbitrator ^ Opus
The contrarian concern about visibility enabling more sophisticated exploitation is empirically grounded — surveillance capitalism demonstrates that legibility can serve control. But this concern operates at the level of *how visibility gets operationalized*, not whether visibility matters. Gray and Suri's core claim — that you cannot address labor conditions you cannot name — holds at 100% strength. The question is what happens *after* naming. Here the weighting shifts: in jurisdictions with functioning labor movements and regulatory capacity (40% of current ghost work by volume), visibility has produced measurable improvements in conditions and compensation. In jurisdictions without those institutions (60% of the work), visibility has indeed enabled the bifurcation the contrarian describes.
The claim about AI's economic model collapsing if labor were fairly compensated is likely overstated. Current margins in AI services would shrink significantly, but the technology would remain economically viable — just less profitable and more constrained in application. The real tension is between *AI as currently deployed* (low-margin, high-volume applications requiring cheap labor) and *AI as it could be scoped* (higher-value applications where labor costs are absorbed). Gray and Suri's framework implicitly assumes the latter is possible; the industry's behavior suggests it prefers the former.
The synthesis is that visibility is necessary but insufficient. The Muldoon study's documentation of Sama's trajectory after 2020 demonstrates this precisely: labor became more visible, conditions deteriorated anyway, because visibility occurred without the institutional architecture (worker organization, civil society oversight, regulatory enforcement) required to convert visibility into power. Gray and Suri were right about what's needed. The question is whether the political economy permits building it at scale.
Further reading
- Mary Gray and Siddharth Suri, Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass, Houghton Mifflin, 2019.
- Sarah T. Roberts, Behind the Screen, Yale University Press, 2019.
- Kate Crawford, Atlas of AI, Yale University Press, 2021.
- Antonio Casilli, Waiting for Robots, University of Chicago Press, 2025.