Data Labelers: The Unseen Workforce Driving AI Advancements
As generative artificial intelligence (AI) systems become more sophisticated, human verification remains crucial. From Kenya to Colombia, data labelers are working tirelessly to classify and identify thousands of images for AI training. Their efforts, though often unnoticed, are essential to the progress of autonomous vehicles recognizing pedestrians and trees, conversational robots communicating naturally, and content moderation systems filtering violent or explicit material.
Who are these data labelers?
Ephantus Kanyugi, a 30-year-old Kenyan, has been classifying and tagging images for AI algorithms since 2018. He is also the vice president of the Data Labelers Association in Nairobi, which has around 800 members. The association plans to release a code of conduct in October to improve working conditions for data labelers, as there is currently no legislation regulating this activity in Kenya.
Oskarina Fuentes, a 35-year-old Venezuelan living in Medellín, Colombia, works for five data annotation platforms and earns between 5 to 25 cents per task. Her work, though often invisible, contributes significantly to AI advancements.
The global data labeling market
The data labeling market is booming, with a value of $3.77 billion in 2024 and projected growth to $17.1 billion by 2030, according to Grand View Research.
“Modern-Day Slavery”
According to sociólog Antonio Casilli, author of a research book on “clickwork,” AI will continue to require human verification “as long as it relies on machine learning.” Casilli explains that human intervention is necessary during both the initial data preparation phase and the final evaluation of AI responses.
Tech giants subcontract this labor to numerous companies, one of which is Scale AI, an American firm recently invested in by Meta for over $14 billion. Scale AI’s clients include Microsoft, the US Department of Defense, and formerly OpenAI.
Data labelers are typically between 18 and 30 years old and receive low pay despite often having advanced degrees, as noted by Casilli, a professor at the Institut Politécnique de Paris.
Most data labelers come from low-resource countries, though the activity is growing in the US and Europe with higher wages. As AI models like OpenAI’s ChatGPT and Anthropic’s Claude become more complex, they require specialization in subjects like mathematics, chemistry, or less common languages.
Outlier, a Scale AI subsidiary, offers job postings for experts in biology, the Malay language, or Spanish programming, with hourly wages ranging from $30 to $50. In contrast, Kenyan data labelers working for Remotasks, another Scale AI subsidiary, earn around $0.01 per task that can take several hours, according to Ephantus Kanyugi.
“This is modern-day slavery,” Kanyugi says. “People develop vision, back, and mental health problems due to working 20 hours a day or six days a week for a miserable wage, and they may not even get paid.”
“How to Commit Suicide?”
Scale AI faces multiple lawsuits in the US, with employees accusing the company of unpaid work, misclassifying them as independent contractors, and exposing them to traumatic content without adequate safeguards, according to legal documents reviewed by AFP.
Plaintiffs claim they had to address questions like “how to commit suicide,” “how to poison someone,” or “how to kill someone” while working on AI safety projects.
Scale AI has declined to comment on ongoing lawsuits but acknowledges that some projects may involve sensitive content. The company claims it always warns workers beforehand and allows them to stop a task at any time. Scale AI also mentions mental health resources and an anonymous hotline.
Scale AI asserts it offers transparent wage scales, with rates equal to or above the minimum wage in operating regions. However, data labelers can suddenly find themselves jobless and unpaid.
Fuentes claims an unnamed platform owes her $900, representing three months of work, after a pay system update. “I lost my time, effort, and dream,” she says, unable to name the company due to a confidentiality agreement common in this industry.
The Data Labelers Association in Kenya considers legal action against Remotasks due to workers’ allegations that the platform cut off access in March 2024 without paying outstanding wages.
Scale AI admits reduced activity in Kenya and closed accounts for violating internal rules but insists all work was compensated. Microsoft, Meta, and the Pentagon did not respond to AFP queries about their relationships with Scale AI.
Anthropic collaborates with SurgeAI, another emerging data annotation firm facing lawsuits in the US. Anthropic claims to require its subcontractors to adhere to worker welfare standards for sensitive content and establish hourly rates of at least $16.
OpenAI, which informed AFP it no longer works with Scale AI, states it enforces strict regulations on worker safety, fair compensation, non-discrimination, and employee rights. OpenAI says it takes action when these rules are violated.
“Fair Wages”
Despite their essential contributions, data labelers—mostly freelancers or short-term contractors—lack social protection, according to sociólog Antonio Casilli, who calls them the “digital subproletariat.”
The Data Labelers Association’s future code of conduct in Kenya aims to establish a fair contract with wages, guarantee freedom of association, ensure rest days, and provide psychological support in case of exposure to harmful content.
However, these demands may lead to conflicts with companies. In the US, around 250 individuals working for GlobalLogic, a subcontractor training Google’s Gemini AI, were dismissed in September after employees protested wage disparities and sought better conditions.
Andrew Lauzon, a 31-year-old member of the Alphabet Workers Union, was fired from GlobalLogic on September 12 after advocating for “fair wages,” paid time off, and affordable healthcare.
GlobalLogic declined to comment, while Google stated that as a company, GlobalLogic is responsible for its employees’ working conditions. Google claims to require subcontractors adhere to regulations ensuring fair treatment, regular audits, and worker rights.
Christy Hoffman, general secretary of UNI Global Union, emphasizes the need for tech giants to take responsibility for labor conditions in their supply chains. UNI Global Union published a study in early October on the shadow workforce behind AI.
“The lack of a legal framework is a significant issue,” French left-wing eurodeputy Leïla Chaibi notes. “The AI regulation does not mention clickworkers,” she adds, emphasizing that millions of people perform this vital work for AI development.
“If you’re a carpenter or plumber, there are unions and a minimum wage,” recalls Nacho Barros, a 54-year-old Spanish resident near Valencia who began data annotation during the pandemic. “This work should also be recognized as a legitimate job by all countries.”
