Who Does the Hidden Human Work Behind AI?

An AI assistant can draft a report in seconds. Before that, people may have spent hours marking examples, checking answers, and removing material no product team wants in a demo. A Brown University story about a student mapping the labor behind AI points to a basic fact of the industry: the interface looks automated, but much of the work is not.

The important question is not whether humans help build AI. They do. It is which tasks they perform, under what conditions, and whether the companies selling automated systems make that dependency visible.

What the reporting and research establish

Brown University recently highlighted a student project mapping the global labor that supports AI. The headline captures a genuine problem of visibility: workers can be geographically far from the companies and users who benefit from their work, and their contribution can disappear behind a model’s polished interface. The project is a prompt to examine the labor chain, not a complete census of everyone working on AI.

That chain includes several different jobs. Data workers may label images or classify text so a system can learn patterns. Evaluators compare model responses against instructions or rate which answer is more useful. Content moderators review disturbing material used to identify unsafe outputs. Other workers transcribe speech, check data quality, or maintain the infrastructure that makes these tasks possible.

The International Labour Organization’s 2021 report on digital labor platforms examined online and location-based platform work across 21 countries. Its research included surveys of about 12,000 workers and interviews with 85 businesses. The ILO reported that the number of digital labor platforms had grown from at least 142 in 2010 to more than 777 in 2020. Those figures cover platform work broadly, not AI data workers alone. They show the scale of the infrastructure through which companies can source tasks, rather than the number of people training AI.

Researchers Mary L. Gray and Siddharth Suri described this wider arrangement in Ghost Work, documenting how digital services rely on human workers whose tasks are often hidden from users. Their work predates the current generative-AI boom, but the underlying arrangement is familiar: software presents a seamless result while people handle exceptions, judgments, and cleanup behind the scenes.

Interpretation: automation can move work out of sight

Calling a system automated can be accurate about how a particular output is produced and misleading about the full process. A chatbot may generate a reply without a person typing that reply. But people may have helped select its training examples, judged earlier responses, or reviewed reports of harmful behavior. “Automated” describes the moment of use; it does not describe the whole supply chain.

This distinction matters because the work is not all interchangeable clicking. Some tasks are repetitive, but others require language skills, local context, or the ability to recognize a subtle threat. A reviewer deciding whether a response encourages self-harm, for example, is doing a different job from someone tagging a picture of a stop sign. Grouping both under “data labeling” makes the work easier to overlook and harder to assess.

The business incentives can reinforce that invisibility. Buyers want large volumes of labeled examples and quick turnaround. Vendors compete to supply them. Workers may be hired through subcontractors or online platforms, making it harder for an end customer to know who performed a task, what they were paid, or what support they received. That does not prove every contract is abusive. It does mean that a company’s polished product page is poor evidence about labor conditions.

Our earlier piece on CAPTCHAs and human verification work looks at a related trade: people and platforms can be part of the machinery that appears to distinguish humans from bots. The connection is not that every CAPTCHA trains an AI model. It is that digital systems often depend on human judgment while presenting the result as a simple technical operation.

Why better models may still need people

More capable models can reduce demand for some forms of routine labeling. They can also create new review work. A model that writes fluently can produce mistakes that are harder to spot than obvious nonsense. Businesses still need ways to test whether answers are accurate, safe, and appropriate for particular users.

That creates a measurement problem. A benchmark score can summarize performance on a test set, but it cannot by itself establish that a system behaves well in every setting. Human review helps identify failures that were not anticipated when the benchmark was designed. This is one reason the debate over AI evaluation is not just a contest between better models and cheaper computing. It is also about who defines a good answer and who is assigned to check it.

Readers may recognize a similar gap between visible software and its maintenance in the rules that govern digital services. In both cases, the interface hides decisions that shape what users can do. With AI labor, the hidden decision is often who bears the cost of making a system seem effortless.

Prediction: disclosure will become a procurement question

As organizations buy AI tools, they are likely to ask vendors more specific questions about data sourcing, evaluation, and human review. That pressure may come from regulators, institutional buyers, or customers who need to understand how a system was tested. It may also produce more public reporting, though companies will have incentives to disclose favorable facts and omit inconvenient detail.

Disclosure alone will not settle whether a job is fair. A useful account would distinguish direct employees from contractors, name the kinds of work involved, explain how workers can challenge a decision, and describe protections for exposure to disturbing material. It would also avoid presenting a single pay figure without context about task availability, unpaid time, and subcontracting.

There is a risk that better automation simply pushes human work further down the chain, where it is more fragmented and harder to audit. There is also a chance that stronger contracts and clearer standards improve conditions. Neither outcome follows automatically from a more capable model. Procurement rules and labor protections will do more to determine the result than the model’s fluency.

The useful way to read an AI claim

When a company says a system can perform a task, ask what happened before and after the demonstration. Who prepared the examples? Who checked the result? Who handles the cases the model gets wrong? The answers do not make the technology less impressive. They make its actual capabilities, costs, and limits easier to judge.

AI’s human labor is not a footnote to the product. It is part of the product’s operating conditions. Making that work visible is the first step toward deciding whether the system is genuinely efficient, or merely efficient at moving effort somewhere the customer cannot see.