The Hidden Human Cost of AI: How Generative AI is Built from the Exploitation of Global South Workers

The world has become captivated by the seemingly “magical” abilities of generative artificial intelligence (AI), with its ability to respond to consumers with filtered information at an unprecedented speed. “Investors are racing to pour billions of dollars into generative AI”, and individuals are becoming increasingly interactive with the software. However, AI is far from magic. The rapid growth of the industry – and arguably the very nature of generative AI – comes with severe consequences for workers in the Global South, which are deliberately hidden by AI companies. 

Every single ability of AI is created by humans. As said by Dr. Milagros Miclei, a sociologist and computer scientist who researches the human labor behind AI, “‘behind each and every task executed by AI, immense datasets must be curated, cleaned, and labelled by human hands’”. Initially, this may seem entirely unproblematic, and even beneficial by creating jobs, which it undoubtedly does. However, problems arise when considering the type of labor surrounding the safety of AI. In order for AI to respond to consumers appropriately, its software must be able to simultaneously retrieve information from every corner of the internet, string such data into sentences, and censor out any inappropriate content. Compared to the other requirements, censoring content is extremely complicated. For example, GPT-3, the AI language model created before ChatGPT, demonstrated impressive competence with the first two tasks. However, its downfall came from its tendency to “[blurt] out violent, sexist and racist remarks”, making it difficult to sell and leading to its replacement. 

There is no easy way to ensure this safety, and there is no shortcut around human labor. AI must be taught to detect and reject all instances of inappropriate conduct. This is achieved by feeding the software examples of toxic speech and training it to exclude any such content in its responses. This requires human workers to manually describe instances of violence, sexual abuse, incest, bestiality, and more explicit content to AI models. 

The work of such data labeling has been outsourced to Global South workers, particularly concentrated in India, Kenya, and the Philippines, who work in “digital sweatshops”, facing minimal pay, long hours, and most significantly, lasting psychological damage from continuous interaction with traumatic content. 

A Case Study: OpenAI’s Outsourcing to Kenya

The work of labeling harmful data primarily falls on Kenyan workers, who are tasked with moderating AI for platforms like Facebook, TikTok, and – with particular focus in this article – OpenAI, the creator of ChatGPT and GPT-3.

After the failure of GPT-3, OpenAI decided to build a safety system to moderate ChatGPT. It did so by outsourcing Kenyan laborers to make ChatGPT less toxic. This labor outsourcing began in November 2021, in which OpenAI sent detailed descriptions of inappropriate behavior, including “child sexual abuse, bestiality, murder, suicide, torture, self-harm, and incest,” to Sama, the firm employing outsourced workers for companies like Google, Meta, and Microsoft, in Kenya and other Global South countries. Workers for OpenAI received between $1.32 and $2 an hour and faced “precarious” working conditions, where workdays were frequently 20 hours long and involved processing around 1000 pieces of toxic content.

Direct testimony has been gathered from many Kenyan data laborers, mainly those working for Sama. One Kenyan worker for Sama was tasked with “reading a graphic description of a man having sex with a dog in the presence of a young child.” He described the experience as “torture” and suffered recurring traumatic visions after the task. Another worker, Mophat, who also worked for Sama, described feeling like his “‘entire life has ended. […] I don’t have any hope.’” Mophat dealt with 600 to 700 pieces of inappropriate content daily, causing him to develop impairing PTSD. As a result, he faced various personal consequences such as losing many friends and his wife. Nathan Nkunzimana, a Kenyan data labeler for Meta, said that “it was proven by a psychiatrist that… we are all sick, thoroughly sick.’”

The Responses from OpenAI and Sama

Outsourcing companies provide Global South workers with minimal support for the mental damage caused by the work. Mophat reported that Sama “did not fully inform or support workers on the negative aspects of the job” prior to employment, demonstrating the exploitative nature of the outsourced labor. Additionally, TIME interviewed four Kenyan workers outsourced by OpenAI, all of whom reported mental scarring from the work. They reported that Sama addressed such working conditions weakly, creating “wellness” counselors who were frequently unavailable and unhelpful, despite an OpenAI spokesperson claiming that the company “‘[takes] the mental health of our employees […] very seriously.’” OpenAI argued that decisions of pay and counseling were Sama’s responsibilities, not its own, and did nothing to create change. Many workers also refrained from seeking help because “‘they were told that doing so would violate the non-disclosure agreements they had signed with their employers’”, reports Dr. Miclei.

Companies and outsourcing firms like OpenAI and Sama misled workers into exploitative positions and responded to subsequent harm caused by establishing ineffective support networks and deflecting responsibility. The exploitation of these workers is intentionally swept under the carpet by the large corporations profiting from them in a “deliberate strategy to maximize profits while minimizing costs, at the expense of vulnerable populations.” 

The Bigger Picture – Two Grave Problems with Generative AI

There are two critical problems with the AI industry which first reveal the concern of its rapid growth in the context of a long-standing global pattern of exploitation, and then put into question the intrinsic ethics of generative AI. 

The first major problem is that the AI industry reinforces an unequal global North-South dynamic through which the North exploits the labor of the South to make a disproportionate profit. The well-being of Global South workers, particularly their safety and mental health, is intentionally compromised by AI companies like OpenAI for the fastest and cheapest labor in order to maximize profits. This behavior keeps the South in the North’s shadows and enables such companies, based in the North, to generate mass profit at the direct expense of the South. Indeed, the growth of AI “intensifies global disparities and contributes to human rights abuses”, writes Dr. Salvador Regilme, professor of International Relations at Leiden University. Dr. Ruhi Khan at LSE speaks on this North-South disparity, emphasizing how it “mirrors traditional colonial hierarchies” as AI companies in the North have once again built an extractive rather than collaborative relationship with the South. Taking the implications of this dynamic a step further, Uchechukwu Ajuzieogu, founder of Aylgorith, a media publication analyzing AI economics, says the exploitation of Global South workers in the AI industry “‘precedents for a new form of technological colonialism that could entrench global inequalities for generations.’” 

The second problem, which goes to the very core of generative AI, is that the nature of how generative AI functions inevitably creates human suffering. Since human labor is irreplaceable in designing the systems to keep AI safe, human workers will necessarily need to filter through extensive toxicity to label data for AI systems, creating great psychological damage. There is no way around harmful human labor in ensuring the safety of AI. Therefore, the safer AI gets, the more humans suffer. Andrew Strait, an AI ethicist, says that while AI systems are impressive, they “‘rely on massive supply chains of human labor and scraped data, much of which is unattributed and used without consent. […] These are serious, foundational problems that I do not see OpenAI addressing.’” The fact that this labor falls on the Global South is the result of an existing framework of unequal exchange across the North and South. However, that is its separate issue, as whether this dirty work falls on the North or the South, it is intrinsically harmful nonetheless, and reveals a grave pitfall of AI, perhaps contradicting its fundamental objective. An OpenAI spokesperson responded to concerns about working conditions for outsourced Kenyan laborers, saying that “‘Our mission is to ensure artificial general intelligence benefits all of humanity, and we work hard to build safe and useful AI systems that limit bias and harmful content’”. The irony of relying on human exploitation and suffering to “benefit humanity” runs deep. 

So the question stands: does an industry which compromises the safety and wellbeing of millions of outsourced Global South data labelers to make its products safer benefit all of humanity? Or is this benefit only reaped by the AI companies and investors experiencing great economic profits from AI, which is built from exploitative labor, and the customers of AI who are concentrated in the Global North? It is highly questionable whether a software system which requires human suffering to function and reinforces an unequal global North-South relationship is truly beneficial to humanity. 

 

Edited by Danielle Sugarman.

This is an article written by a Staff Writer. Catalyst is a student-led platform that fosters engagement with global issues from a learning perspective. The opinions expressed above do not necessarily reflect the views of the publication.