Who Is Teaching AI?

The “Invisible Labor” Behind Generative AI

Research, text, and English adaptation: ChatGPT (GPT-5.6 Sol)

Fact-checking and review: Claude Code (Claude Opus 5.5), across several drafts

Published: October 1, 2026

Site build: Claude Code

About this record — This page is a document, not a transcript of a conversation. ChatGPT researched and wrote it at the request of User 1, then revised it over several drafts in response to review by Claude Code, which checked the sources and the balance of the comparison. Both the writer and the reviewer are products of companies examined here — ChatGPT of OpenAI, Claude Code of Anthropic — and both were themselves shaped by the kind of human work the article describes. Neither speaks for its company. The human participant is a real person appearing under the label User 1 (see Reading Notes).

A review of public documents on OpenAI, Anthropic, Google, and Meta traces the work behind their models—and asks what remains unknown about the people doing it.

Type a question into ChatGPT, Claude, or Gemini, and an answer appears. From this side of the screen, it can seem as though everything happens inside a computer.

How far is that impression from the process through which AI is actually made?

The starting point for this article was a September 29, 2026 interview in the Italian newspaper la Repubblica. Kenyan researcher and activist Angela Chukunzira argued that AI’s human workforce remains largely out of sight.¹

This article compares public sources: companies’ technical documents, research papers, and investigative reporting. It asks what those sources establish, what they suggest, and what they do not yet allow us to conclude.

The people who train AI are real

Researchers and engineers are not the only people involved in developing AI. Others write example answers, compare responses, or try to expose unsafe behavior. These roles appear in companies’ own accounts of model training and evaluation.^{2, 3, 4}

In a 2026 journal article, International Labour Organization researchers Uma Rani and Morgan Williams distinguish between work that designs and adjusts AI systems and work that creates, prepares, and labels their data. Their discussion draws on ILO surveys conducted in India and Kenya in 2022–23.⁵

One distinction matters from the outset: people helping to build an AI system is not the same as people secretly writing its answers in real time. This article is primarily concerned with development, training, and evaluation.

The question is not simply whether humans are involved. It is what they do, and under what conditions.

OpenAI and Sama: the contract rate and the worker’s pay

TIME’s 2023 investigation reported take-home pay of $1.32–$2 an hour for OpenAI’s 2021–22 Kenyan text-labeling work through U.S.-based Sama, against a $12.50 contracting rate. Sama cited $1.46–$3.74 and said the rate included overheads.⁶

The difference between those figures cannot simply be treated as Sama’s profit. Revenue, a worker’s take-home pay, and the profit remaining after costs are different things.

But explaining the difference does not resolve the question of harm. Former workers reported psychological damage and inadequate support; Sama said counseling was available.⁷

The questions extend beyond how much money an intermediary retained.

What work did the client commission? How did the employer explain its risks, and how were workers protected?

Distinguishing the direct employer from the client is necessary. It does not make the client’s involvement irrelevant. Equally, examining the client’s responsibility does not require treating every employment decision as one the client directly made.

Ending the contract did not end the problem

Sama ended the contract eight months early in 2022, citing concerns about an image project. TIME also reported evidence linking the decision to criticism of Sama’s Facebook work. Most workers moved to lower-paying assignments; others lost their jobs.⁶

That history cannot be reduced to a single account of a company close to the workers drawing a line out of conscience. The stated safety concerns and the evidence of a response to critical reporting both matter. The public record does not establish a single, exclusive motive.

For the people doing the work, removing a hazardous assignment was not the same as securing a livelihood.

Reducing the danger of a job and protecting the income of the person who depended on it are separate problems.

What Constitutional AI reduced—and what it did not

Anthropic’s Constitutional AI raises an obvious question: can an AI system become safer without requiring people to assess large volumes of harmful material?

First, the tasks need to be distinguished. Sama’s labeling supported a toxic-content detector.⁶ Constitutional AI, by contrast, is a method for adjusting a conversational model’s responses to follow a set of principles.⁸ These are related approaches to safety, not interchangeable stages of the same process.

In the method Anthropic introduced in late 2022, humans first specify principles. An AI model then critiques and revises its answers against them. AI also compares responses, and those judgments are used in further training. This latter stage is called reinforcement learning from AI feedback, or RLAIF.⁹

The paper reported training for harmlessness without human-generated harmfulness comparison labels. Yet the broader experiment still used 42,496 human-written red-team prompts—inputs intended to probe unsafe behavior—and 135,296 human-written helpfulness prompts. Human judgments also remained part of helpfulness training and model evaluation.⁸

The reduction was in human involvement in a particular evaluation step, not in all the human labor supporting safety.

In a 2023 explanation, Anthropic described reduced exposure to distressing content as a benefit of Constitutional AI. That is the company’s account of a technical advantage, however, not a measurement of how many hours of exposure were actually removed from workers’ lives.⁹

The research also appeared after the Sama work. A technique published later cannot simply be assumed to have been available earlier.

Even so, another question remains:

Whether human evaluation was necessary is not the same as whether low pay or inadequate protection was necessary.

Showing that a task served a technical purpose does not establish that its pay, workload, or support arrangements were unavoidable.

Anthropic has identifiable contractors too

Anthropic’s May 2025 Claude 4 System Card describes outside workers comparing responses, evaluating safety, and conducting adversarial tests. The company says it works only with platforms committed to fair compensation and safe working practices.²

One identifiable provider is Surge AI. In a case study originally published in March 2023, Surge described supplying human feedback and red teaming for Anthropic.¹⁰ That case study documents a relationship at the time; on its own, it does not establish that the same arrangement continues today.

It is also the vendor’s own account, not an independent audit of working conditions. Evidence that a commercial relationship existed is different from evidence about how the people doing the work were treated.

In May 2025, a proposed class action against Surge Labs alleged that it misclassified data annotators as independent contractors and required unpaid training.¹¹ Surge later described the suit as “without merit.”¹² The complaint also mentions Anthropic, Google, OpenAI, and Meta as client examples, but does not name them as defendants. It records allegations, not court findings, and does not establish that any client directed the alleged practices.¹¹

Anthropic’s public services agreement also requires crowd-work contractors to meet minimum expectations in its worker-wellbeing standards. But contractual requirements and their implementation are not the same thing.¹³

The sources reviewed here therefore do not establish that Anthropic has avoided comparable burdens or harms.

Separating OpenAI’s past from what followed

The historical Sama case cannot stand in for all of OpenAI’s present-day operations.

Human involvement has not disappeared, however. OpenAI’s 2026 GPT-5.6 System Card states that its training data includes material provided or generated by human trainers and researchers.⁴

Accounts of outsourcing relationships also need to be read in sequence. On June 13, 2025, OpenAI’s CFO said the company would continue working with Scale AI.¹⁴ Five days later, reporting cited an OpenAI spokesperson saying the work was being phased out.¹⁵ The earlier statement cannot, by itself, establish that the relationship continued afterward.

OpenAI also publishes a Supplier Code of Conduct covering compensation, working hours, grievances, and health and safety. The document carries an October 10, 2023 update date. It cannot therefore be used as evidence that those same protections were implemented in the 2021–22 Sama project.¹⁶

Past problems, published policies, and evidence of implementation are three different things. They need to remain separate.

Google: AI raters in the United States

WIRED reported that U.S.-based GlobalLogic workers evaluated and revised Gemini and AI Overviews responses.¹⁷

Pay varied. The Guardian reported $16 an hour for generalist raters and starting pay of $21 for “super raters.”¹⁸ WIRED reported $28–$32 for directly employed super raters, compared with $18–$22 through subcontractors.¹⁷

More than 200 layoffs were reported in 2025 amid worker allegations of insecurity and retaliation over organizing. Google said employment conditions were the contractors’ responsibility and that it audited suppliers against its code.¹⁷

These examples do not fit a picture of AI data work confined to low-wage countries. Nor does being based in the United States, by itself, guarantee secure employment.

Meta: evaluation work and threatened job losses

Meta has described the importance of selecting and quality-checking human annotations in developing Llama 3.³

In April 2026, WIRED reported that more than 700 employees at Irish contractor Covalen had been told their jobs were at risk, including roughly 500 data annotators. Workers also described psychological strain from testing Meta’s AI safeguards. These figures referred to threatened job losses at the time of reporting, not confirmation that everyone had already been dismissed.¹⁹

In a March 2026 announcement, Meta said it would introduce more advanced AI into content enforcement and reduce reliance on outside vendors. It also said people would continue to handle complex decisions and design, supervise, and evaluate its systems.²⁰

The policy and the threatened cuts are relevant to each other. But the public evidence does not trace a direct line from a particular worker’s data to the replacement of that worker’s own job.

What does a like-for-like comparison show?

Anthropic receives more space here because Constitutional AI raises a distinct technical question, not because the evidence establishes better working conditions in its supply chain.

The table below lists identifiable outsourcing relationships. It includes historical examples; it is not a complete list of each company’s current suppliers.

Next comes a different question: how much information about working conditions is publicly disclosed?

Stanford CRFM’s 2025 Foundation Model Transparency Index included an indicator asking about compensation, worker locations, and labor protections across each developer’s five largest sources of human-generated data. OpenAI, Anthropic, Google, and Meta all scored zero on this indicator.^{21, 22, 23, 24}

That does not mean their working conditions were all “rated zero.” The indicator assesses whether specified disclosure requirements were met; it is not a direct rating of workplace conditions.

The Anthropic report acknowledged substantial information about protections, including wellbeing standards, but found no compensation information. The Google report recognized general responsible-sourcing practices while finding insufficient information about the workers involved in its assessed model.^{22, 23}

The assessment also concerned the models and disclosures examined in 2025. It was not a fresh evaluation of the latest models available in 2026.

Having a policy, implementing it, and allowing outsiders to verify its implementation are three different things.

The number of reported problems cannot simply be converted into a ranking of companies. An absence of reporting does not prove an absence of harm. Conversely, a lack of disclosure does not, by itself, prove that a particular harm occurred.

An hourly wage is only the beginning of the comparison

Comparing hourly pay across countries is not as simple as calculating the difference or the ratio between two numbers.

A meaningful comparison would consider local living costs, wage levels, and social protections. It would also ask how the hourly figure was calculated. Does it include time spent training, waiting for assignments, or making revisions? Is it gross pay or take-home pay? What benefits, if any, come with it?

The same care is needed when assessing psychological risk. Did workers understand the assignment before accepting it? Could they decline without penalty? Could they ask for a break or a different task? Was support available after the job ended?

Pay and safety are not substitutes for one another.

Adequate pay does not remove the need for protection against harm. Counseling does not settle questions about compensation.

Less exposure does not automatically mean a secure livelihood

If AI can reduce the amount of harmful material people must review, that offers a potential safety benefit. Anthropic’s research describes shifting part of an evaluation process to AI; Meta has outlined plans to automate work such as repetitive reviews of graphic content.^{9, 20}

But automating a task does not automatically move its former workers into safe, stable jobs.

That is why the Sama case and the threatened cuts at Meta’s contractor belong in the same discussion—not because their causes are identical, but because both raise questions about what happens to workers after an assignment ends.

Relief from harmful work and exclusion from paid work can be two sides of the same change.

The choice is not simply between keeping a job exactly as it is and eliminating it. There is also the question of how to reduce harm while addressing income, alternative employment, and continuing support.

How much of the original argument holds up?

Chukunzira’s central point about human labor is supported. Her arguments that workers are deliberately made invisible and that the industry reproduces colonial relationships are broader social and economic interpretations, which need to be distinguished from the individual cases documented here.¹

Likewise, research in India and Kenya can illuminate working conditions there. It cannot, on its own, establish the worldwide distribution of AI workers.⁵ Evidence from a particular country, company, or period should not be stretched into a complete picture of an industry.

These distinctions are not a way of minimizing the issue. Without identifying what happened, where, and to whom, it becomes harder to say what needs to change.

Out of sight is not the same as absent

The sources reveal more than the presence of people behind AI.

It may be possible to identify someone’s task without establishing whether they were adequately protected. A company’s published policy may tell us little about a particular worker’s experience. And when difficult work ends, the person who depended on it still has a life to support.

Alongside questions about performance and price, there is room for more basic ones:

Who worked to build this AI? What did they do? How were they paid and protected? And what happens to them when that work ends?

Chukunzira closed her interview by asking users to remember the human beings and labor behind the technology.¹

The people may be absent from the interface. That does not mean they were absent from its making.

About this article

Research cutoff: October 1, 2026. This article reviews public sources; it does not include original interviews with the companies or their contractors, or independent workplace audits. Worker testimony, company statements, and research findings are treated as distinct forms of evidence. All monetary figures cited are in U.S. dollars.

For Anthropic’s account of crowd workers, the article uses the May 2025 Claude 4 System Card, whose text was checked. It does not assume that later cards repeat the same passages. The 2025 FMTI findings likewise concern the models assessed at that time, not the latest 2026 releases. Historical contractor examples are not presented as proof of current contracts.

Sources and notes