Andrew Zeng, Stanford University
Once a high-performing master’s student studying oil engineering at the height of Venezuela’s oil boom, Oskarina Fuentes Anaya had been forced to become a refugee in Colombia by the middle of the 2010s, when oil prices plummeted and Venezuela was plunged into an economic crisis. Originally working in a call center, Fuentes confronted further catastrophe when she was diagnosed with acute diabetes, leaving her with few remaining career options. In this environment of desperation, Fuentes turned to working for Appen, an AI company which paid her a few cents for every task she completed.
At first, Fuentes made enough money performing menial content moderation and product categorization to support herself. But soon, the troubles began with Appen. First, as more and more refugees came to Colombia, Appen stepped up its recruitment efforts—and with its generally safer platform than companies like Hive Micro, which often showed gorey and unsettling imagery, it attracted hundreds of thousands of workers. But soon Fuentes found her task list drying up: She began turning on notifications so that she could claim tasks in the middle of the night. And in the midst of all of this, Appen wrongfully suspended her account, and it took months of pleading through Google Translate to get it reinstated.
Despite all of the difficulties, however, Fuentes considers herself relatively lucky. “I’ve survived because of this platform,” she told the MIT Technology Review as part of a 2022 series on AI colonialism. “Other platforms have stopped paying, but Appen has always been there.” Indeed, the emergence of AI platforms like Appen provided refugees with a lifeline toward securing a stable income and recovering from economic catastrophe. But hidden beneath this lies a much deeper problem: Periods of economic crisis are of direct benefit to AI companies seeking cheap labor to train their models.
The cost of recruiting a tester for a large language model (LLM) in the United States is incredibly high. It’s certainly higher than the cost of recruiting a tester from a job-starved country from the Global South. So in the search for scalability and profits, frontier AI companies have been recruiting testers from third-world countries, where pay is low and regulations are lax. The results for their employees have often been heartwrenching.
In January 2023, TIME published an article revealing that in its drive to acquire human feedback at a cheap price, OpenAI used the San Francisco-based firm Sama to outsource work to Kenyan workers. The workers, who were paid between $1.32 and $2 per hour, sifted through graphic text depicting bestiality, child sexual abuse, and murder. A Guardian follow-up quoted Mophat Okinyi, one of the Kenyan workers, who stated that he “lost [his] family” after he grew extremely paranoid from reading texts about rapists, leading his pregnant wife to call him a changed man and leave. Workers testified that no mental health services were provided by Sama, which pulled out of its collaboration with OpenAI 8 months early, leaving the workers scarred in its wake.
Similarly, a WIRED article in September 2023 revealed that Finnish prisoners were employed by Metroc, a Finnish startup, to train its language models. The prisoners made roughly $1.67 an hour, with the prison stating that its partnership with the AI firm was meant to prepare prisoners for the digital age. Some observers, however, were far from convinced. In an interview published as part of the article, Amos Toh, a senior researcher on AI at Human Rights Watch, pointed out that while AI is often placed on a pedestal for its potential to bring about a utopian automated society, it is important to remember that “there are actual human people powering a lot of these systems.”
The list of AI companies looking to third-world countries in search of cheap labor goes on, and on, and on. Some cases, like that of the Finnish prisoners, seem to be rather benign; others, like that of the Kenyan workers, seem much more dangerous.
One could reasonably argue from an economic standpoint that the employment of sweatshop workers helped drive China’s economic growth—some evidence suggests that it even fueled economic development across the globe. But while these arguments are in many ways suspect, the point of this piece is not to argue against this framework. Instead, this article only points out that because human feedback is in such high demand for the purpose of training LLMs, artificial intelligence companies have an incentive to keep some countries poor in the long run so that their citizens can continue to provide cheap labor. Furthermore, as different AI companies and start-ups compete to improve their programs, it is very reasonable that there will be a race to the bottom in wages, and that without regulation, mental health resources will continue to be denied to poor workers who sift through large quantities of deeply disturbing data.
The three examples I’ve mentioned above—Venezuelan refugees, Kenyan workers, and Finnish prisoners—all have one thing in common, along with many other testers that have been hired by large AI companies: They are unable to effectively bargain for better wages and working conditions. These conditions provide the ideal environment for low-cost work.
In the future, perhaps the work of interacting with disturbing content will be left to artificial intelligence systems themselves. But until that day comes, it is crucial that effective mechanisms be put in place to hold frontier AI companies accountable for the damage done by work they foist onto victims of economic crises. It is likewise critical that AI companies are further regulated so that other documented forms of exploitation—including wrongful terminations and the withholding of wages—be expunged along with the perverse incentives that make such work possible.