Invisible Labor Behind AI: Who Keeps the Systems Running

Behind every AI-generated text, every automated image classification, and every autonomous vehicle lies human labor—performed by millions of people whose names, faces, and working conditions are systematically kept invisible. This article reveals who actually operates the AI infrastructure, what numbers define the market, and what consequences arise for the use of AI in content production.

When a language model drafts a marketing campaign or an image generator creates a visual, the output appears machine-made and autonomous. The reality is different: every one of these systems was trained, corrected, and supervised by humans. Data annotation, Reinforcement Learning from Human Feedback (RLHF), and content moderation are not peripheral phenomena—they form the foundation of every AI workflow. According to the Stanford HAI AI Index Report 2025, industry produces over 90% of all frontier models, and their productivity gains are built on precisely this human contribution.¹

For anyone deploying AI tools in AI content production, this raises an uncomfortable question: under what conditions is the infrastructure built that underpins your marketing processes? The answer touches on ethics, reputation, and—very concretely—budget planning for the years ahead. The data basis of this article draws on six current sources from 2025 and 2026: the GIZ/Aapti report on working conditions in the Global South, a peer-reviewed analysis from AI & Society, the Stanford HAI Index, two market forecasts from Oxford Economics and Research and Markets, and a study by AWU-CWA on US-based data workers.²³⁴⁵⁶

AI systems are not autonomous: the concept of heteromation

No commercial AI system operates without continuous human input. The concept of heteromation—coined by researchers Ekbia and Nardi and further developed by Nemer and Sobral (2025)—describes how platforms accumulate capital by extracting cheap, computer-mediated human labor and marketing it as machine performance.³ The work disappears behind interfaces and APIs, while the value creation remains with the platform operators.

Self-driving vehicles and their human remote control

Autonomous vehicles require hundreds of thousands of so-called contractors who draw bounding boxes around objects, classify street scenes, and intervene remotely in real time when the system reaches its limits. According to the GIZ report, individual providers employ over 100,000 external workers for these tasks.² The vehicles do not drive autonomously—they drive with an invisible workforce in the background.

Language models and the human feedback behind them

Large Language Models (LLMs) undergo a phase of RLHF after pre-training. Humans evaluate responses, flag toxic content, and prioritize helpful outputs. Workers in Kenya received less than 2 USD per hour for this work—which included reviewing content depicting violence, abuse, and extremism.² Without this labor, no language model would be market-ready.

Note: Every AI tool you use in your AI workflow relies on human data annotation. The question is not whether, but under what conditions this work takes place.

Market size of invisible labor: a billion-dollar industry

The human labor behind AI systems is not a niche phenomenon—it is a global industry with double-digit growth rates. Two independent market analyses from 2025 quantify the volume and dynamics of this sector—and both show: the costs for AI infrastructure will rise, not fall.

Data annotation: 8 to 15 billion USD in revenue

Oxford Economics estimates the global data annotation market for 2024 at 8.2 to 15.1 billion USD.⁵ The range is explained by different scope definitions: some estimates capture only pure labeling services, while others include quality assurance and data preparation. Even the conservative lower bound exceeds the annual revenue of many DAX-listed companies.

Microtasking market: tripling by 2030

The microtasking market—the outsourcing of granular data tasks via platforms—is growing according to Research and Markets from 5.1 billion USD (2025) to 14.6 billion USD (2030) at a compound annual growth rate of 23.4%.⁴ This growth is driven directly by rising demand for AI training data. For budget owners, this means: AI tool prices will increase in the medium term because input costs are increasing.

Expert tip: When budgeting for AI-powered content production, factor in an annual price increase of 15–20%. Data annotation costs are driving tool prices upward.

Demographics of data workers: young, educated, underpaid

The people behind AI infrastructure are not unskilled laborers. They are predominantly between 20 and 35 years old, hold university degrees, and live in the Global South—in India, the Philippines, Kenya, or Venezuela.² Their qualifications stand in stark contrast to their pay and their legal employment status.

Pay gap for identical work

For the same annotation task, US-based workers receive approximately 15 USD per hour, while their counterparts in Kenya receive less than 2 USD per hour.²⁶ The pay gap is a factor of seven to eight—for identical work, identical output, and identical quality requirements. Platforms deliberately exploit wage differentials between countries to maximize their margins.

No labor protections due to contractor status

Classification as "Independent Contractors" removes these workers from the labor law protections of their respective countries. No termination protection, no health insurance, no paid leave, no pension entitlements. The GIZ report documents that platforms can suspend accounts without warning—meaning a person's entire income depends on an algorithm.²

Psychological costs of AI content production

The invisible labor behind AI systems causes documented health damage. Content moderators are particularly affected—those who must review and classify toxic content so that language models and social media platforms can function safely. The consequences range from sleep disorders to suicide attempts.²

12-hour shifts with traumatizing content

Content moderators work shifts of up to twelve hours, reviewing depictions of violence, sexual abuse, and child pornography. The GIZ/Aapti report documents cases of post-traumatic stress disorder (PTSD), chronic depression, and sleep disorders among these workers.² Psychological support is either not provided at all or only minimally offered by the platforms.

Non-disclosure clauses prevent access to help

Non-Disclosure Agreements (NDAs) prohibit workers from discussing their work and its effects—including with therapists or family members. These clauses create a double isolation: the work itself is traumatizing, and access to support is contractually blocked.² Lawsuits against this practice—in Kenya, for example—have been ongoing since 2023 without fundamentally changing industry standards.

Expert tip: Ask your AI tool providers explicitly about their content moderation policies and the working conditions of the people involved. Transparency starts with specific questions.

Structural invisibility as a business model

The invisibility of human labor behind AI is not an oversight—it is a design principle. Platforms, subcontractor chains, and algorithmic management work together to maintain the impression of full automation. For companies using AI tools, this means: the supply chain is intentionally designed to be opaque.

Human labor framed as machine performance

The term "Artificial Artificial Intelligence"—originally coined by a major microtasking platform—captures the principle precisely: human labor is packaged to look like an API response.³ Nemer and Sobral (2025) demonstrate in their analysis that this framing is not a marketing gimmick but a systematic strategy for the devaluation of human labor. When work is invisible, it becomes easier to underpay.

Algorithmic management and real-time surveillance

Data workers are subject to a level of surveillance that would be unthinkable in traditional employment relationships. Keystroke tracking, real-time productivity measurement, and automatic penalties for deviating from time targets are standard.² Those who work too slowly lose assignments. Those who work too fast are suspended for quality issues. The corridor is narrow, and the decision is made by an algorithm with no right of appeal.

Multi-layered subcontractor chains obscure accountability

Between the AI company that publishes a model and the person who annotates data, there are three to five layers of subcontractors. Each layer reduces transparency and dilutes accountability. The GIZ report describes cases where workers do not even know which end product they are labeling data for.² This structure makes due diligence extremely difficult for AI tool buyers—but not impossible.

AI does not eliminate work: it displaces it

The widespread assumption that "AI replaces human work" is false in this sweeping generality. AI displaces work—geographically to the Global South, legally into precarious contractor arrangements, and visually behind platform interfaces. The Stanford HAI Index 2025 confirms that productivity gains from AI are built on human labor.¹ New jobs are created—but under conditions that contradict the progress narrative.

A current example of this hybridization: a major mobility provider has since 2025 been having drivers in twelve Indian cities label data on the side—while they wait for passengers.² The boundary between "gig work" and "data work" is blurring. For companies purchasing AI tools, this makes the supply chain even more opaque. The exact number of data workers involved worldwide is unknown—estimates range from several million to the tens of millions. The opacity of supply chains prevents precise quantification.

Consequences for the use of AI in marketing

Every AI tool in content production is built on the supply chain described above. This is not an abstract ethical problem—it is a concrete reputational risk and a due diligence issue. The EU Platform Work Directive 2024/2831 will increase regulatory pressure on these structures. Those who ask the right questions today will avoid unpleasant surprises tomorrow.

Actively demand supply chain transparency

Ask your AI tool providers: Who annotates the training data? Under what contractual conditions? In which countries? At what compensation? Most providers will not answer these questions proactively—but the question itself sends a signal. Companies like Crispy Content® already integrate such assessments into their evaluation of AI tools for client projects, because ethical supply chains are part of consistent brand communication.

Ethical AI use as a competitive differentiator

Brands that transparently communicate how they use AI and what standards they apply to their supply chain build trust. This is especially true in B2B contexts, where business partners increasingly apply ESG criteria to their service providers. Ethical AI use is not a feel-good topic—it is a competitive factor in tenders and partnership negotiations.

Budget planning: factor in rising costs

The supposedly low-cost AI content production externalizes costs onto precarious workers. When regulatory pressure and rising data annotation costs (CAGR 23.4%) make this externalization more expensive, tool prices will increase. Do not plan your AI budgets based on today's prices—instead, anticipate annual increases of 15–25% over the next three to five years.

1. Create an AI tool inventory: List all AI tools active in your marketing workflow.
2. Query providers: Request information from each provider on data annotation, working conditions, and subcontractor chains.
3. Conduct a risk assessment: Evaluate each tool for reputational risk, regulatory risk, and pricing risk.
4. Identify alternatives: Determine which providers can demonstrate transparent supply chains.
5. Adapt your communication strategy: Integrate your AI usage standards into external brand communication.

Design hybrid workflows deliberately instead of concealing them

"AI-generated content" is never purely machine-made. This insight changes how quality and accountability are assessed in content strategy. Those who deliberately design hybrid workflows—the combination of human expertise and AI support—and communicate them transparently position themselves more credibly than companies that conceal the human contribution.

The parallel to your own organization is obvious: if global AI companies make the human labor behind their products invisible, marketing teams should take the opposite approach. State clearly where AI assists and where humans decide. This builds trust with customers, partners, and your own workforce. An AI workflow that makes human expertise visible is more resilient in the long run than one that simulates full automation.

Note: The EU Platform Work Directive 2024/2831 will be implemented incrementally from 2026 onward. Companies using AI tools with opaque supply chains face a growing regulatory risk.

Frequently asked questions

What does invisible labor mean in the context of AI infrastructure?

Invisible labor refers to the human work—data annotation, RLHF, content moderation—that makes AI systems functional but is systematically hidden by platforms and providers. Several million people worldwide perform this work under precarious conditions.

How large is the market for data annotation and microtasking?

The global data annotation market comprises 8.2 to 15.1 billion USD in revenue (2024). The microtasking market is growing from 5.1 billion USD (2025) to a projected 14.6 billion USD (2030)—a compound annual growth rate of 23.4%.

What risks arise for companies using AI tools in their AI workflow?

Three risk categories: reputational risk through association with exploitative working conditions, regulatory risk from the EU Platform Work Directive 2024/2831, and pricing risk from rising data annotation costs that will increase AI tool prices in the medium term.

How can marketing teams audit the supply chain of their AI content production?

Ask your AI tool providers specifically about annotation locations, compensation models, subcontractor structures, and the legal employment status of data workers. Document the responses and evaluate each tool by degree of transparency and risk potential.

Will AI tools become more expensive due to rising labor costs?

Yes. Data annotation costs are growing at a CAGR of 23.4%. Additionally, EU regulation is increasing pressure for fair working conditions. Both factors will raise the operating costs of AI providers and be reflected in tool prices—likely with annual increases of 15–25% over the next three to five years.

Sources:

¹ Stanford University (2025): AI Index Report 2025. URL: https://hai.stanford.edu/ai-index/2025-ai-index-report (accessed May 25, 2026).

² GIZ / Aapti Institute (2026): Invisible Workers, Visible Harms: Perils and Precarities of AI Labour. URL: https://www.bmz-digital.global/wp-content/uploads/2026/02/GIZ_2026_InvisibleWorkersVisibleHarms-1.pdf (accessed May 25, 2026).

³ Nemer, D. / Sobral, V. (2025): Artificial Intelligence as Heteromation: The Human Infrastructure Behind the Machine. In: AI & Society, Springer. URL: https://link.springer.com/article/10.1007/s00146-025-02664-5 (accessed May 25, 2026).

⁴ Research and Markets (2025): Microtasking Market Forecasts Report 2025–2030. URL: https://uk.finance.yahoo.com/news/microtasking-market-forecasts-report-2025-091300646.html (accessed May 25, 2026).

⁵ Oxford Economics / Scale AI (2025): Impact of Data Annotation. URL: https://static.scale.com/uploads/6691558a94899f2f65a87a75/Oxford%20Economics_%20Impact%20of%20Data%20Annotation%20%20Dec%202025.pdf (accessed May 25, 2026).

⁶ AWU-CWA / TechEquity (2025): Ghost Workers in the AI Machine. URL: https://cwa-union.org/ghost-workers-ai-machine (accessed May 25, 2026).