Ghosts in the AI Machine – In a nutshell

From voice assistants to tenant screening tools, AI systems have entered many facets of everyday life. More often than not, “artificial intelligence” models are presented as fully automated, having dispensed with the need for human intervention.

In reality, AI systems are built on the essential labor of tens of thousands of data workers in the U.S. and globally, who train, test, and maintain them. Big tech companies in the U.S. and Europe subcontract data work through long chains of suppliers and online platforms. Research examining global AI supply chains has revealed how data workers in the global South, including Kenya, India, Venezuela, the Philippines, and elsewhere, perform the human labor behind purportedly automated systems, often for low wages under traumatizing conditions. But less is known about the working conditions of U.S.-based data workers.

Conventional wisdom suggests that these domestic workers thrive in well-paid jobs, while employers seek to reduce costs by outsourcing work to regions with lower wages and fewer labor regulations. However, by combining data from a survey of 160 data workers with insights from 15 in-depth interviews, TechEquity and the Alphabet Workers Union-CWA found that poor working conditions, commonly observed in the Global South, are also prevalent in data work in the U.S.—even as the AI industry is booming. 

Here’s an overview of our main findings from our report, “Ghost Workers in the AI Machine: U.S. Data Workers Speak Out About Big Tech’s Exploitation.”

What do data workers do?

The workers surveyed for this report have a wide range of job titles, and workers’ descriptions of their day-to-day responsibilities provide insight into how central their work is to functional AI systems. For example, some of the workers who participated in our research described their roles as:

• Search quality rater – evaluating Google search results for relevance and overall page quality;
• Content writer – writing technical explainers for users; 
• LLM evaluator – checking model outputs for factual accuracy, tone, grammar, punctuation, and relevance; 
• Internet safety evaluator – evaluating content such as YouTube videos to ensure they follow safety guidelines.

Who are US data workers?

Workers are employed by a web of vendors and subcontractors, making it difficult for workers to address workplace problems. In our survey, respondents reported being employed by 28 individual vendors, with the largest number employed by the vendors Telus, GlobalLogic, Scale AI, and WeLocalize.

On top of that, 23% of respondents said they work as 1099 independent contractors for vendors like Scale AI. 1099 independent contractors lack the basic legal rights of a directly-employed (W-2) employee, including minimum wage, overtime, workers’ compensation insurance, and protection from anti-discrimination and disability rights laws.

What are the working conditions for the production of AI?

Workers struggle to make ends meet with poorly paid, insecure work

Across the entire survey sample, workers reported a median hourly wage of $15.23 and a median workweek of 29 hours of paid time, which equates to annual earnings of $22,620. Only 23% of respondents are covered by health insurance from their employer.

As a result of these low wages and inadequate benefits, 86% of surveyed workers worry about meeting their financial responsibilities often (63%) or sometimes (26%). 25% of respondents rely on public assistance, primarily food assistance and Medicaid. A shocking 22% percent of respondents report they have experienced homelessness because they could not afford to move or find housing that they could afford.

I do not feel fairly compensated. The pay level is abysmal for the amount of work we put in. I set my schedule but it’s a big dedication: I work 8 hrs a day, it’s just from home. I’m only getting $14/hr before tax so that’s about $12/hr after tax: that’s not sustainable where I live in Richmond, Virginia and it’s my main source of income…

Senior Rater for Bravens working on Google projects

Accessing this poorly paid work comes at a cost for workers. Most workers surveyed (85%) are expected to be on-call even for periods when there are no tasks available to complete. More than half of respondents (66%) report spending at least three hours weekly sitting at their computers waiting for tasks to be available, and 26% report spending more than eight hours waiting for tasks. Only 30% of respondents reported that they are paid for the time when no tasks are available.

In addition to the unpaid time spent waiting for tasks, 26% of respondents reported they were not paid for training time. Even more commonly, workers are not paid for the time it takes to test to qualify for jobs, with 53% of respondents reporting they are not paid for testing.

Even when workers do find tasks, they remain subject to suspension by the automated systems used by AI vendors, which flag workers who fail to meet metrics that are not always transparent to workers and may prevent workers from accessing the platform. More than a quarter (31%) of respondents reported they have experienced being suspended, of whom 88% were not paid during the suspension period. 

I would also like transparency on the platform end because we are never told what happened when a suspension happened. They say they don’t reveal much in order to prevent scammers, but you don’t punish the good people to punish the bad people: we deserve to be told why we are losing our livelihood.

Krista Pawloski Michigan-based worker at Amazon MTurk

Workers lack control over their work

Workers took pride in their work in shaping AI outputs, but often felt that unclear employer metrics prevented them from doing a good job, and expressed frustration at being stuck in dead-end jobs with little scope for career progression or advancement. 

When asked if the metrics to remain in good standing at this job were clearly stated by the employer, most survey respondents said the metrics are somewhat unclear (50%) or very unclear (36%), while only 14% felt that the metrics are completely clear.

Workers’ control over their work process is also undermined by unrealistic time estimates—also known as “average estimated times” (AETs)—to finish tasks and a lack of guidelines and training. Of the 69% of survey respondents whose job requires them to meet predetermined AETs for handling tasks, more than half felt that AETs are often not long enough to complete the task accurately. In interviews, some workers also indicated that exceeding AETs on tasks can lead to work being unpaid.

Compounding this, most respondents report they are regularly assigned tasks for which they are not adequately trained, with only 13% reporting this never happens. 

Workers receive few benefits

Workers frequently deal with a range of violent, offensive, and other sensitive content, and 70% of survey respondents report that their employers asked them to sign a consent form for sensitive content. While a majority of workers said that they were allowed to opt out of sensitive content, many spoke about the fear of not having any tasks available or being let go if they did opt out. 

I have complex PTSD, among other things, so it usually causes several problems for my well-being. The worst part, however, is knowing that I have to stay silent and not complain/continue working, because if not, I could be let go at any moment.

Anonymous Digitive worker from Texas

Although it is now widely acknowledged that exposure to such content can cause significant mental harm, 62% of workers said their employer provides no mental health support. Thirty-four percent of survey respondents reported having a disability, but only 7% (three out of 42) reported having an accommodation at their job.

Despite the many problems outlined above, many workers said that they value the remote nature and flexible hours of data jobs, which is a reason many want to stay in their positions – if they can improve the pay and conditions. Workers spoke in interviews about how they got involved in data work as a way to earn an income after they struggled to find other flexible employment that worked for them due to health issues, mental or physical disabilities, or care responsibilities.

What are the risks of AI systems?

On the frontlines of the development of AI systems, data workers are also well placed to identify the risks of these tools. A majority (52%) of surveyed workers believe they are training AI to replace other workers’ jobs, and 36% believe they are training AI to replace their own jobs. 

Workers describe training models to create art and graphics, and do tasks related to editing, software development, customer service, and employee recruitment, among others. People working on training and testing these tools also brought up concerns about the safety and accuracy of AI systems.

Working on a Google project, a GlobalLogic worker said, “I wish the public knew how imperfect these models were. I wish they knew how unreliable they were, and I also wish they knew just how much slack we as AI workers are pulling to not embarrass these companies, and they still manage to embarrass themselves… In my opinion, they shouldn’t be rolling these things out as things people can use. It’s not ready. It’s really imperfect… I would tell people, if you’re going to use AI, double-check everything they tell you.”

What must be done?

In our worker survey and interviews, data workers explained how their work contributes to the AI industry and how better working conditions could improve the technology itself. Data workers should be treated as skilled workers with expertise to enhance this industry. However, workers face significant structural barriers to having their voices heard.

Data workers are already organizing for change in their industry by unionizing with Alphabet Workers Union-CWA. For example, raters at GlobalLogic, a subsidiary of the multinational conglomerate Hitachi, are fighting for standardized pay and fair raises, ‍equitable paid time off benefits for all workers, job security, and clear employment pathways—despite alleging retaliation for speaking out.

What can we do?

If you want to support these efforts, follow unions and advocacy organizations like Alphabet Workers Union-CWA and Turkopticon to get the latest in research, updates on organizing efforts, and ways to help.

Check out “Ghost Workers in the AI Machine: U.S. Data Workers Speak Out About Big Tech’s Exploitation” to learn more.
If you’re a data worker or have any information on the data work industry, please consider sharing your story and/or first-hand knowledge of data work companies. By sharing your experience, you can help lawmakers understand why they should support bills and regulations to protect everyday people.