The Invisible Labor Force Powering AI

The conversational prowess of AI chatbots like ChatGPT, Gemini, and Claude appears to stem from sophisticated algorithms alone, but this apparent autonomy masks a critical dependency: after these AI models are created, there is a constant, large-scale infusion of human judgment to get them ready for prime time.

To make these models safe, coherent, and aligned with human values, they must be trained and refined by people. This process involves tasks like generating high-quality training examples, rating different model responses to reinforce good behavior, and flagging toxic, biased, or harmful content.

This essential work is generally not done by the high-paid engineers at OpenAI, Google, and Anthropic. Instead, it is performed by a vast, largely invisible labor force in the form of millions of gig workers, primarily in countries like Kenya, Uganda, the Philippines, and India. These workers are often paid as little as $2 per hour by intermediary firms like Scale AI, Surge AI, and Sama, which supply the armies of human raters that AI labs depend upon for scale and cost efficiency.

As a result, the world’s most advanced technology depends on cheap, precarious human labor. This model is coming under increasing strain. With reports of worker trauma from flagging violent content, class-action lawsuits over wage theft, and labor regulator investigations, the fragile human infrastructure that props up AI is beginning to crack, forcing a reckoning over how AI is truly built.

Essential, Permanent Infrastructure

A low-cost labor force is essential to how today’s AI models function. Human workers are needed at every stage of AI production for tasks like creating and annotating data, reinforcing models, and moderating content.

“Today’s frontier models are not self-made. They’re socio-technical systems whose quality and safety hinge on human labor,” said Mark Graham, a professor at the University of Oxford Internet Institute and a director of the Fairwork project, which evaluates digital labor platforms. In his book Feeding the Machine: the Hidden Human Labor Powering AI (Bloomsbury, 2024), Graham and his co-authors illustrate that this global workforce is essential to making these systems usable.

“Without an ongoing, large human-in-the-loop layer, current capabilities would be far more brittle and misaligned, especially on safety-critical or culturally sensitive tasks,” Graham said.

Antonio Casilli, a professor of sociology at the Institute Polytechnique de Paris and author of the book Waiting for Robots: The Hired Hands of Automation (University of Chicago Press, 2025), is similarly direct. “Human-provided data is absolutely essential. There’s no getting around it. These models couldn’t exist without it,” he said.

Kai-Hsin Hung, a researcher at business school HEC Montréal, concurred, noting that human judgment is indispensable. Only humans, he said, can evaluate model outputs for factuality, social nuance, and safety, especially on long-tail edge cases.

“Without continuous human feedback, the model’s behavior drifts, safety deteriorates, and subtle harms remain undetected,” Hung said.

Why the Gig Model?

The industry’s reliance on a distributed, gig-work model goes back years. Hung points to the creation of the ImageNet database around 2007 as the moment that set the referential data practices and work organization for modern AI training.

The ImageNet team calculated it would take undergraduate students 19 years to annotate the millions of images required. They instead turned to Amazon Mechanical Turk, scaling to 50,000 workers across 167 countries, paying them about $0.02 per annotation.

“It was driven by efficiency and cost,” Hung said.

Today, cost remains a primary motivator.

“Let’s be clear: money is the primary driving force here,” Casilli said. “An in-house model is impractical primarily because it would cost vastly more.”

However, cost is not the only factor. Graham noted that cost arbitrage plays a role, but it is not the whole explanation. AI labs, he said, need extreme scale and elasticity, meaning millions of small, episodic tasks that can be staffed up or down at short notice, as well as broad linguistic and cultural coverage that no single in-house team can reproduce.

Casilli pointed to a more strategic benefit: deniability. By classifying workers as remote subcontractors rather than employees, AI companies perpetuate the myth of fully automated, self-sufficient AI solutions, a narrative that appeals to investors.

“The legal distance created by outsourcing labor is a deliberate strategy for AI companies to avoid accountability and political scrutiny,” Casilli said.

Cracks in the Model

This strategy of outsourcing and denial is now facing intense public and legal scrutiny.

Sama’s Kenyan contractors described being traumatized while labeling violent and sexual content for OpenAI. Surge AI is facing a class-action lawsuit over alleged wage theft, while Scale AI is reportedly being investigated by labor regulators. And in the summer of 2025, hundreds of contractors working on Google’s Gemini were laid off after raising concerns over pay.

“We’re already seeing cracks in this model,” Casilli said. “They are mainly due to workers organizing, in the U.S., Europe, Africa, Asia.” This is coupled with shifting public opinion and increasing pressure from regulators.

What happens if this global labor supply is significantly disrupted?

Graham described the core challenge of regulation. This work exists in a labor market where tasks move across borders, platforms, and layers of subcontracting in minutes. “That mobility makes unilateral regulation difficult, because once one jurisdiction tightens standards, the work can quickly flow elsewhere,” he said.

For these reasons, Graham said, the most credible levers are transnational due-diligence regimes, such as the European Union’s Corporate Sustainability Due Diligence Directive.

Graham’s team at Fairwork has also established the Fairwork AI Certification to audit suppliers. “We work with lead firms in AI supply chains to audit their suppliers against the Fairwork principles and to use their purchasing power to raise standards globally,” he said.

Hung agrees accountability is moving up the value chain. Initiatives like the EU’s AI Act and its proposed Corporate Sustainability Due Diligence Directive (CSDDD), an EU law requiring large companies to identify, prevent, and mitigate adverse human rights and environmental impacts across their global value chains, extend responsibility across the entire AI value chain. This, he said, will make AI firms accountable to adopt more responsible data work practices and policies for the workers who train their models.

Casilli said this disruption could be an opportunity. If companies were forced to employ workers directly and pay fairly, he said, “The cost of producing training data and running reinforcement learning would multiply severalfold.”

While AI labs may claim this would collapse the industry, Casilli said it would create “genuine incentives to invest in alternative business models and new technological paradigms.”

Logan Kugler is a technology writer specializing in artificial intelligence based in Tampa, FL, USA. He has been a regular contributor to CACM for 15 years and has written for nearly 100 major publications.

You Just Read

The Invisible Labor Force Powering AI

© 2026 ACM 0001-0782/26/2