Redistributed

General

The Hidden Labor of AI

Published March 8, 2026

laborghost-workredistributionwork

AI doesn't replace human labor. It obscures it.

Behind every "automated" system is a supply chain of human workers: data labelers in Nairobi, content moderators in Manila, RLHF trainers across the globe. The labor didn't disappear — it was redistributed to places where it's cheaper and less visible. Understanding this supply chain is essential to understanding How Language Models Actually Work — and to seeing through the Autonomy Illusion that surrounds AI marketing.

The Automation Myth

Mary Gray and Siddharth Suri call this "ghost work" — the human labor that powers services advertised as AI (Gray & Suri, 2019). Amazon's Mechanical Turk was named after an 18th-century chess-playing "automaton" that hid a human operator inside. The metaphor is exact.

The scale is staggering. Kate Crawford (2021) documents how ImageNet, a foundational dataset for computer vision, was labeled by nearly 50,000 Mechanical Turk workers at roughly $2 per hour. The models trained on this labor are worth billions; the workers who made them possible earned subsistence wages.

This isn't a bug in the system. It's the business model.

Data Labeling: The Foundation

Every supervised machine learning system requires labeled training data. Someone has to look at images and click "cat" or "not cat," read text and mark "toxic" or "safe," watch videos and annotate "pedestrian at 3 seconds, cyclist at 7."

Who does this work?

Scale AI, one of the largest data labeling companies, contracts workers through vendors in Kenya, the Philippines, Venezuela, and elsewhere. Payments average $1-2 per hour — legal in those jurisdictions, but far below what the AI companies' engineers earn (Perrigo, 2023).

Remotasks, another major platform, structures work as microtasks: label this image for $0.03, transcribe this audio clip for $0.05. Workers report earning $1-3 per hour after accounting for unpaid time spent finding and qualifying for tasks (Hao, 2022).

The conditions:

  • No employment relationship — workers are "independent contractors," with no benefits, no job security, no recourse
  • Algorithmic management — platforms rate workers, allocate tasks, and terminate accounts based on opaque metrics
  • Wage theft — rejected work isn't paid, even when the rejection is arbitrary
  • Psychological burden — many tasks involve disturbing content (violence, abuse, hate speech)

Sama, formerly Samasource, positions itself as an "ethical AI" company providing dignified employment in East Africa. Investigations revealed workers earning $1.50/hour to label traumatic content for OpenAI's ChatGPT, with inadequate mental health support (Perrigo, 2023). The full picture of data labeling in the Global South and the RLHF Workforce reveals this as systemic, not isolated.

Not all companies have chosen this path. Karya, a nonprofit platform in India, pays workers 20 times the prevailing rate and shares data licensing profits with annotators. The model proves that dignified data labeling is economically viable — it just requires treating labor as a contribution to be fairly compensated rather than a cost to be minimized.

Content Moderation: The Invisible Shield

Every platform with user-generated content employs content moderators — humans who review posts flagged by algorithms or users and decide what stays up.

Sarah Roberts (2019) documented the toll in Behind the Screen. Moderators spend eight-hour shifts watching videos of violence, sexual abuse, self-harm, and worse. The work is traumatic by design: someone has to see the worst of the internet so you don't have to.

The redistribution:

Meta employs roughly 15,000 content moderators, mostly through contractors like Accenture and Majorel. They're concentrated in countries with lower labor costs: the Philippines, India, Morocco, Kenya. The workers who protect American users from graphic violence earn a fraction of American wages (Newton, 2019).

Why this matters for AI:

Content moderation data feeds back into the Training Pipeline. When moderators label a post as "hate speech," that label trains the next generation of automated classifiers. The trauma is compounded — first in viewing, then in perpetuating a system that will inevitably make mistakes on edge cases. The full scale of the mental health crisis in this workforce deserves dedicated attention.

RLHF: Shaping Model Behavior

Large language models like ChatGPT aren't just trained on text — they're refined through Reinforcement Learning from Human Feedback (RLHF). Human workers read model outputs and rate them for helpfulness, harmlessness, and accuracy.

This is the most direct form of ghost work in generative AI. The model's "personality" — its tendency to be cautious about medical advice, its refusal to generate violent content, its writing style — is shaped by thousands of human judgments.

Who are the RLHF workers?

OpenAI contracted Sama to provide workers in Kenya for "data labeling services." The workers described the task: read AI outputs describing child abuse, torture, incest, and other disturbing content, then label them so the model would learn not to generate such material (Perrigo, 2023).

Pay: $1.32 to $2 per hour. Workers reported lasting trauma. The contract was terminated early, but the model retained their labor.

The exchange:

  • Users get a "safe" chatbot that won't produce disturbing content
  • OpenAI gets a product worth billions
  • Workers in Kenya get $2/hour and psychological damage

This is redistribution — but in the wrong direction.

Algorithmic Management: Workers Serving the Algorithm

The other side of AI labor isn't making the systems — it's working under them.

Kellogg, Valentine, and Christin (2020) identify six mechanisms of algorithmic control:

  1. Directing — algorithms assign tasks, routes, schedules
  2. Evaluating — platforms rate workers based on metrics they can't see or contest
  3. Disciplining — low ratings mean fewer assignments or termination
  4. Recording — every action is logged and analyzed
  5. Restricting — workers can't see the full picture of how decisions are made
  6. Replacing — the constant threat that automation will eliminate the job entirely

Amazon warehouses epitomize algorithmic management. Workers are directed by handheld devices that track their location, measure their "rate" (items picked per hour), and flag deviations. Time off task — even bathroom breaks — is monitored. Workers report feeling like "robots" serving the algorithm (Guendelsberger, 2019).

Uber and Lyft use algorithms to set prices, allocate rides, and evaluate drivers — but the algorithms are opaque. Drivers can't see how surge pricing works, why they're getting fewer rides, or what factors determine their rating. They're managed by a black box (Rosenblat, 2018).

Call centers increasingly use AI to monitor calls, evaluate "sentiment," and coach workers in real time. The software tells workers to smile (it detects vocal tone), to speed up (the call is taking too long), or to use specific phrases. Autonomy disappears (Ajunwa, 2020).

What Workers Are Demanding

Labor organizing in AI is nascent but growing.

Unions in tech:

The Alphabet Workers Union (Google) includes both engineers and contract workers. Amazon's JFK8 warehouse in Staten Island voted to unionize in 2022 — the first US Amazon warehouse to do so.

Content moderator organizing:

In Kenya, Daniel Motaung led a group of former Sama moderators to sue the company and Meta over labor conditions and psychological harm. The case is ongoing (Mureithi, 2022).

RLHF worker demands:

The African Content Moderators Union, formed in 2023, advocates for fair pay, mental health support, and the right to organize across borders.

Regulatory approaches:

  • The EU AI Act requires documentation of human oversight in high-risk systems — but doesn't specify conditions for those humans
  • California's AB5 attempted to classify gig workers as employees, with partial success
  • The EU Platform Work Directive (2024) creates presumption of employment for platform workers

What Redistribution Would Look Like

The problem isn't that humans are involved in AI — it's how that labor is valued and protected.

Fair compensation: If AI companies are worth hundreds of billions, the workers who make their products possible deserve more than $2/hour. Some proposals: compute-weighted royalties, profit-sharing, minimum per-task rates.

Employment status: Ghost workers should be employees, not contractors. This means benefits, job security, and the right to organize.

Mental health: Content moderation and RLHF work is traumatic. Companies should provide robust mental health support, limit exposure, and pay premiums for hazardous work.

Transparency: Workers should understand how they're evaluated and have recourse when they're treated unfairly. Algorithmic management needs accountability.

Global standards: Labor arbitrage only works because standards differ across countries. International coordination — through trade agreements, ILO standards, or cross-border organizing — could raise the floor.

The value AI creates doesn't come from nowhere. It comes from labor — often hidden, often poorly paid, often concentrated in the Global South. Redistribution means making that labor visible and ensuring workers share in the value they create.

The good news: visibility is increasing. The African Content Moderators Union, lawsuits in Kenya and the Philippines, investigative journalism by TIME and The Verge, and academic work by Crawford, Gray, and Roberts have made it harder for the industry to treat this labor as invisible. The collective bargaining movement and policy frameworks like the EU AI Act create institutional pressure for transparency. What remains is translating visibility into structural change — from awareness to accountability to actual redistribution of the value these workers create.

Sources

  • Ajunwa, Ifeoma. "The Paradox of Automation as Anti-Bias Intervention." Cardozo Law Review, 2020. SSRN
  • Crawford, Kate. Atlas of AI. Yale University Press, 2021. Publisher
  • Gray, Mary L. and Siddharth Suri. Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt, 2019. Publisher
  • Guendelsberger, Emily. On the Clock: What Low-Wage Work Did to Me and How It Drives America Insane. Little, Brown, 2019. Publisher
  • Hao, Karen. "Artificial Intelligence Is Built on the Backs of Underpaid Workers." MIT Technology Review, 2022. Link
  • Irani, Lilly. "Difference and Dependence Among Digital Workers: The Case of Amazon Mechanical Turk." South Atlantic Quarterly, 2015. Duke
  • Kellogg, Katherine C., Melissa A. Valentine, and Angèle Christin. "Algorithms at Work: The New Contested Terrain of Control." Academy of Management Annals, 2020. AoM
  • Mureithi, Carlos. "Kenyan Workers Sue Meta Over Alleged Exploitation." Rest of World, 2022. Link
  • Newton, Casey. "The Trauma Floor: The Secret Lives of Facebook Moderators in America." The Verge, 2019. Link
  • Perrigo, Billy. "OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic." TIME, 2023. Link
  • Roberts, Sarah T. Behind the Screen: Content Moderation in the Shadows of Social Media. Yale University Press, 2019. Publisher
  • Rosenblat, Alex. Uberland: How Algorithms Are Rewriting the Rules of Work. University of California Press, 2018. Publisher
  • Tubaro, Paola, Antonio A. Casilli, and Marion Coville. "The Trainer, the Verifier, the Imitator: Three Ways in Which Human Platform Workers Support Artificial Intelligence." Big Data & Society, 2020. Sage
← Back to all articles
On this page