The need for organization and mutual support among workers is vital to improving their working conditions and increasing bargaining power and improving the quality of the work.
Lílian quit her formal job last year. Working outside the home had become a “nightmare” since she started looking after her young daughter. Through a TikTok video, Lílian learned about the possibility of working online training artificial intelligence.
After a month of attempts, Lílian passed the tests authorizing her to work at Appen, a platform that subcontracts workers to meet the growing demand from big tech companies for the production, classification, and analysis of data.
Today, she works six days a week on a flexible schedule to “improve artificial intelligence with data,” as Appen advertises. If all goes well, she earns about R$ 1,400 ($250) monthly, without any other benefits.
Lílian belongs to a class of workers often referred to as ghosts, hide-aways, or micro-workers. Through multinational platforms like Tellus, OneForma, and Appen, big tech companies hire mass-scale cheap labor to perform rote tasks worldwide.
Tech giants like Meta, Google, and TikTok profit from the ease of purchasing datasets from workers whose asking price is far below that of comparable industry professionals, benefitting from a labor pool with few legal protections. Shrouded in a veil of mystery by opaque confidentiality agreements, these big tech companies create an environment where people often fail to realize who or what entities they work for.
In addition to receiving low wages, outsourced workers undergo limited training and work under tight deadlines. There are numerous reports of unpaid salaries, contracts unilaterally terminated without explanation, and a lack of support from the platforms.
According to Rafael Grohmann-a professor at the University of Toronto and a researcher at the Fairwork network advocating for decent work principles on platforms-the model harkens back to 19th-century discussions about industrial capitalism. “We’re quite vintage, having to revisit theories we thought were outdated,” he says.
Outsourcing is a strategy that avoids connecting workers to their final client, says Milagros Miceli, sociologist, computer scientist, and founder of a research group on algorithms and ethics at the Weizenbaum Institute in Berlin.
“It’s important for the workers themselves to realize they’re contributing to a multibillion-dollar industry,” she said. “If you realize you’re working for Microsoft or Google, you’ll ask for more money. I think all workers should know who they are working for and the profit generated by their work.”
Brazil is already one of the largest markets for big tech companies recruiting cheap digital labor. Researchers point out that the pandemic intensified this process in recent years due to Brazil’s economic crisis and the desire to stay at home, combined with increased social media and internet consumption.
Today, Brazilians rank fourth in visits to Appen and fifth in visits to Oneforma, behind the US, UK, India, and the Philippines, according to SimilarWeb data.
In Brazil, these platforms operate in a political and legal limbo. There isn’t reliable data on how many Brazilians work for these companies to reference for labor regulation or protective policies.These businesses also don’t have legal representation in Brazil, undermining any possible claim or legal action against them. Meanwhile, domestic AI regulation initiatives have ignored a central issue by not mentioning data workers in a recent regulatory proposal on discussion in the Senate.
Machine learning systems are a form of artificial intelligence that use algorithms to process data and generate outputs. A legion of workers serves as raw material for automation, enabling the so-called intelligent computing systems to learn and mimic patterns.
Without a mountain of content produced by media outlets and real people, ChatGPT wouldn’t be able to provide quality answers. Without real people interpreting typing errors in search results, Google would be unable to guess what you actually meant with that misspelled word. Without workers interpreting photos to train computer vision algorithms, smart cameras wouldn’t be able to identify objects in an image.
Tech companies must hire millions of workers to execute numerous human tasks to develop AI systems, whose products are largely invisible. The cheapest way the industry has found to accomplish this is through intermediary multinational companies.
Thus, workers worldwide are paid to take photos, transcribe audio, give opinions on ads, moderate content, fact-check, identify images, and make purchases. All kinds of tasks are available, including moderators recording videos of themselves dancing, or videos of children. That’s right, children. Appen pays $20 (about R$ 108) for 10 videos like these. The task order provides no information on what will be done with this content.
Intermediary companies brand this type of work as an offer of freedom, as there are no fixed hours and the earnings per task depend on individual performance. However, dozens of internal documents obtained by Intercept Brazil show that these companies exert extreme control over workers.
For example, workers must sign strict confidentiality agreements prohibiting them from exchanging information with other “independent contractors” – as the company calls those who work for it – and from organizing collectively.
The workers aren’t employees of the final clients and, instead, work on projects organized under codenames. In some cases, they’re even prohibited from mentioning them publicly.
“Do not participate in groups or chats outside of the communication methods provided by Appen to discuss confidential information,” says one of the company’s terms of use pages, whose clients include Amazon, Microsoft, Nvidia, and Meta.
The company further requires that work must be done on a device connected to an encrypted network. Sharing any client or project information – including the project’s guidelines and even its name – outside of the company’s internal environment or with individuals who have not signed confidentiality agreements is prohibited.
“Violating any of these legal obligations may result in permanent removal from Appen projects, repercussions for breach of contract, or other legal measures against you,” says one of the terms of use pages.
Appen also instructs that remote work should only be done at home, not in public spaces. Workers are prohibited from providing instructions to one another and are advised to use headphones and screen protectors to prevent others from seeing their work.
OneForma, a platform used for data preparation by Google, also prohibits public mentions of projects. Workers sign an agreement that exempts the company from responsibilities regarding any leaks – meaning that if any client information were to become public, only the worker would be held liable.
OneForma also requires that workers waive any right to collective legal action. “If, for any reason, a claim proceeds in court and not in arbitration, each party waives any right to a jury trial,” note OneForma’s terms of use.
The Intercept sent a series of questions to Appen, OneForma and Tellus. The companies did not answer.
Gabriela, age 36, worked between Appen and Lionbridge for a year and a half. Needing extra cash after renovating her house, Gabriela took the job because she didn’twant to leave her young children alone. “It was a lifesaver. It helped me a lot,” she told Intercept.
Her first task was taking selfies. She sent videos of herself dancing, exercising, chatting, or moving her hands. The material had to meet specific requirements to be accepted, such as proper framing and lighting. Later, Gabriela would also classify ads and purchase products. In her best months, she made up to R$ 800 ($150).
Like most workers, Gabriela violated the company’s rules and joined clandestine support groups. “It helps a lot, you don’t feel alone, people support each other,” she said.
According to Milagros Miceli, who advocates for workers’ decision-making power on tech platforms, the architecture of this market undermines the quality of the data feeding AI systems.
Miceli suggests that data creation may be compromised because of a lack of sense of purpose. Without knowing exactly what they are working for, people typically perform poorly. A second issue with the data is the isolation imposed on workers. “Many studies are showing that if you collaborate, you get better results,” said Miceli.
Assigned to fragmented tasks, workers are alienated from the purpose of their work. They may be contributing to addicting teenagers to repetitive videos or training police surveillance and crime prediction systems without even knowing precisely what they are doing.
“The platforms are not obligated to inform workers about the projects they are working on,” said researcher Phil Jones, author of Work Without the Worker: Labor in the Age of Platform Capitalism.
“The notion of alienation is that of estrangement. In a factory where you only made the wheel of the car, of course, you knew you were in a car manufacturing plant. It’s as if you intensified this estrangement to the point of no longer knowing where the wheel or screw is going,” explained Rafael Grohmann.
Grohmann and Miceli both have evidence of how essential clandestine worker groups are, “not only for the survival of workers in terrible conditions but also for improved performance,” said Miceli. “Workers need to collectivize, organize, and unionize if possible. It’s a collective process that must come from the workers.”
In late December 2023, Marcia received an expected approval: after a full month of trying, she finally passed the test for a remote job at project Uolo. Her role was to evaluate ads on Meta’s networks.
She registered with the subcontractor Appen, who promised a rate of $3.50 per hour worked, almost double the Brazilian minimum wage.
Marcia worked hard. She was supposed to evaluate at least 40 ads per hour, classifying whether they were inappropriate. She says that on an average day, she worked about three hours but only managed to get paid for a little over one hour.
At the end of the first month, she experienced another setback: her computer had a bug and she was unable to work. Marcia asked for tips to buy another notebook in a WhatsApp group. “I only have R$ 500 (~$90),” she said. “It’s going to be difficult,” a colleague replied.
Shortly after, several workers received an email stating that the project was temporarily suspended and access to tasks cut off. “Guys, are we going to be without work for the rest of the week?” a worker asked in the messaging group.
In the job description, the company promised six hours of work per day or 20 hours per week, totaling $700 (R$ 3,780) per month, an attractive paycheck for a Brazilian worker. No one in any of the groups attained that amount. In the clandestine support group, workers waited anxiously for the company to unlock new tasks. “I still need to earn another $260 this month,” said Lílian. “The poor man’s joy is short-lived.”
When they were finally able to start again, workers on the project found out they had suffered a poor evaluation. A low score is reason for concern, as three bad ratings result in suspension.
Lílian’s appointed score was 60%, which was quite low compared to her usual score of 100%. She tried to complain to the platform, but no one responded. Soon, she could no longer log into the platform. She still had $138 (R$ 745) to receive from Appen. “I’ve been without work for three days. My God, someone help me,” she wrote in the workers’ group.
Shortly after, following instructions from colleagues, she reinstalled the VPN necessary for the service and managed to resume. “Appen is leaving me discouraged. I worked two hours on Tuesday and it says I worked six minutes,” she complained.
Feeling worried, Lílian attempted to get onto other projects but failed the qualifying tests. She woke up each morning afraid the project would disappear without explanation. “Today I didn’t receive any emails from Appen, I’m getting worried,” she wrote.
A survey conducted in Brazil in 2020 showed that bugs are common problems for workers on tech platforms. Support groups can help solve these problems, although they are prohibited by the company’s terms of use.
“Communication between workers helps with payment processes, to avoid financial losses,” explain researchers Grohmann and Willian Fernandes Araújo in the article “The factory floor (Brazilian) of artificial intelligence: the production of data and the role of communication between Appen and Lionbridge workers,” which inspired the title of this project.
Two days later, Appen offered two extra hours for the workers. The relief was short-lived. In February, Lílian’s invoice came due with almost $30 missing. Even so, she continues working on the platform, yearning for new tasks.
If a worker is suspended, a queue of people awaits to replace her. “The constant oversupply of work erodes any bargaining power for better payments because for each task there’s a long line of workers willing to get it done for less money,” wrote researcher Florian A. Schmidt, professor of design and media theory at the University of Applied Sciences HTW Dresden, in an article published in 2022.
Schmidt, who studied the online labor market in Venezuela, said that excess demand generates constant anxiety among workers, who incessantly refresh pages to receive new tasks.
“Like the rest of the gig economy, there is little flexibility and autonomy if you have to jump at any opportunity to do a task before others. And this competition is much more extreme if the work is not based on location,” he explained.
Renan Kalil is an attorney at the Ministry of Labor (MPT) who researched the Amazon platform Mechanical Turk for his PhD. According to him, Amazon’s control and authority over the worker offer robust evidence to characterize an employment relationship and justify the application of labor rights.
The problem, he explains, is applying these rights. The MPT has faced difficulties in contacting companies withoutrepresentation in Brazil. “It is not uncommon for the worker to performan activity in one country, the platform to be based in a second country, and the service recipient in a third country.”
“How can the company or the service recipient be cited?” Kalil wonders. “This is one of the factors we’re missing when debating labor regulation.”
In April 2024, the European Union approved a directive to regulate remote work on tech platforms. It provides a presumption of employment if there are mechanisms in place for controlling the worker. It also prohibits dismissals based on decisions made by an algorithm.
The Lula government, which took office in 2023 with a proposal to regulate the issue, simply ignored data workers. The group created for digital platforms, formed by various ministries and labor entities, simply did not discuss the issue in any of its 12 meetings. The focus was on delivery and passenger transport apps. Those working “digitally” were left out.
Brazil is also discussing the regulation of artificial intelligence with legislative proposal 2338, 2023, authored by Senator Rodrigo Pacheco of the PSD party. But the proposal also doesn’t address how databases, the primary sources of these systems, are created, honed, and treated.
“The problem is a step backward. People are not connecting the dots and seeing that we have a labor problem. They only see a problem with data and the system,” said Milagros Miceli.
For Miceli, who participated in a group that took data workers to the European Parliament during their regulation discussions, the structure of these platforms is designed to make workers obey. When the conditions under which data is collected are obscured, the data sets “become a black box.”
Miceli advocates for a change, with more spaces for workers to be heard and have a voice in the process. The issue is also on the radar of the International Labor Organization (ILO). There is still no regulation on the subject but the 113th International Labor Conference, to be held in 2025, will have a session to define international standards on decent working conditions for tech platforms.