AI's "Invisible" Labour Market: From Indian Cashiers to Kenyan Trainers
How much does "invisible labor" from the Global South contribute to big tech's AI success?
Image from the Outsource Accelerator
Hi everyone! I’m so glad to be back from my hiatus. I wanted to take a moment to explain why I have been away. I fell prey to every writer’s nemesis - writer’s block. This time around, however, it was not just an “I can’t put into words what it is I want to say” but also a “How relevant is this anyway?” I read “Do Not Hoard Things” by Patrick Muindi which snapped me out of my self-deprecative spiral. All that is good within us has been created for us to share.
On top of that, I have been thinking about the direction I would like this space to take. Rather than solely being about AI safety, I would like to branch out into futurism and strategic foresight in line with the work I am doing with the wonderful team at Silicone Humanism.
I’m excited for what the next few months in this space will look like and I hope you are as well!
Amazon’s “Just Walk Out”
Roughly eight weeks ago, I opened my X ( formerly Twitter) feed and was met with tens of memes trolling Amazon. It had been revealed to the public that the tech giant’s infamous “Just Walk Out” technology, which had been purported to be a cutting-edge, novel, and ground-breaking sensory technology, was indeed just 1,000 human laborers working in India. The Amazon associates would manually add up the value of goods in a shopper’s cart.
The Just Walk Out technology allows customers to simply walk into a physical Amazon store like Amazon Go or Amazon Fresh, shop, and walk out without having to physically scan their items to pay. Shoppers are debited virtually and receive their receipts via their Amazon account. Amazon states that the technological underpinnings responsible for this innovation are “a combination of computer vision, object recognition, advanced sensors, deep machine learning models, and generative AI.”
While the claim of using Indian manual laborers has since been refuted by Amazon, the truth values of both the allegation and the rebuttal have not been fully assessed. Amazon reports that the Machine Learning (ML) models underlying the Just Walk Out technology are improved iteratively through assessing shopping data that is labeled and annotated by humans. Moreover, Amazon alleges that the associates do not manually keep score and tally the contents of shoppers’ shopping carts. On the other hand, a report states that in 2022, 700 out of every 1000 Just Walk Out sales were manually reviewed by Amazon India’s team.
one of the hilarious tweets that I came across on my X (Twitter) feed
The veracity of the claims aside, we can delve into the nuances of the situation. If it does indeed hold true that Indian cashiers were adding up the value of the baskets and Amazon gave all the credit to the researchers who built the ML models for the Just Walk Out technology, we can make some inferences. First, is that Amazon outsourced the work to remote Indian laborers because that would be cheaper labor on their part, allowing them to pay as little as possible for maximum value receipt. Second, having workers working remotely allows them to be distanced from the true impact of their work, keeping them shielded from the appraisal given to the tech giant’s novel technology.
On Data Annotation and Its Importance
Amazon is not the only tech giant that has adhered to such practices. This situation prompts reflection on broader issues and labor patterns seen not only in the tech industry but the world at large. In my last post, I wrote in detail about the data pre-processing stage where data labeling/ annotation occurs. Data annotation is perhaps one of the most important stages in ML as it sets the precedent for the model’s functioning and continued learning.
Data labeling is described as annotating data to indicate a target that you want a model to predict. The process closely mimics how children are taught to read and make sense of the world. As a child, you were presented with a colorful chart with different letters of the alphabet and corresponding items. Here you learned to make the association between the letter “a”, the word “apple”, and an image of an apple. You did this iteratively until you understood it such that when someone pointed at the picture of an apple on the learning chart, you identified the letter in question as “a”. In ML, meticulous attention to the pre-processing stage is paramount. This step lays the foundation for effective model training, much like how a child's understanding of the world is largely shaped by their foundational knowledge of letters and words.
Sama, Open AI, and Kenyan Moderators
When Open AI was building ChatGPT, they delegated a part of the data labeling task to Sama, a third-party contractor who got workers in Kenya. The specific part of the data labeling that was assigned to the Kenyan contractors was identifying violence, hate speech, and sexual abuse. To successfully do this, the annotators were given descriptive text detailing grotesque scenarios of situations depicting pedophilia, bestiality, murder, suicide, torture, self-harm, and incest. One of the content moderators, Okinyi, gave a detailed account of his experience. He stated that he would go through up to 700 text passages a day. This ultimately made him isolate himself out of paranoia about the scope of evil that humans could do unto one another. The job affected him to the extent that he changed, became a shell of himself, and his expectant wife left him. Okinyi’s story is just one out of the possible thousands that have not been shared.
Other moderators working under Sama in Nairobi revealed that they were ill-prepared for the nature of the content that they would be reviewing. They reported barely receiving any psychological support. Deepening the crisis was the low compensation of between $1.46 and $3.7 per hour remunerated to the workers. For their nine-hour shifts, the moderators were expected to read and annotate between 150-250 passages of text. These portions of text contained anything between 100 to 1000 words. In a statement sent out by OpenAI, the burden of payment and worker well-being was delegated to Sama. The relationship between the two companies collapsed in early 2022. OpenAI had contracted Sama requesting collaboration on a project to collect sexual and violent images. Realizing that some of the images requested were both harmful and illegal, Sama terminated its business relationship with the tech giant; stripping the steady income of hundreds of Kenyan workers. The Sama contractors sued Meta, citing unlawful termination.
A Brief Analysis
While the issues of the Indian Amazon associates and the Kenyan data annotators differ in the details, they are two sides of the same coin. They indicate a much larger issue at play - neocolonial business practices. Outsourcing inexpensive labor from developing countries, while failing to uphold the ethical standards set in place by the company’s domestic country is a form of exploitation. The low wages, unfavorable working conditions, and nature of work would be unacceptable in Western nations. So why choose to delegate the labor to the Global South? The impetus is evident. Throughout history, there has been a lower level of humanity assigned to people from the Global South. A persistent extortion of human capital and natural resources that form the backbone of multi-trillion dollar economies. The labor force of this demographic is seen as comparable to that of a machine - quick, relentless, and devoid of the essential consciousness central to humanity. The people are seen as incapable of experiencing pain both physically and psychologically. This imposed nonchalance can be observed in longstanding industries like the Congo mines and fast fashion sweatshops.
While big tech can be seen as the bad actor within this paradigm, they are not solely to blame. A lack of legal infrastructure to enforce labor laws bears culpability. Practices such as the lack of enforcement of a universal minimum wage in Kenya form fertile breeding ground for unjust business practices. Without a legal framework that seeks to protect its citizens, workers will continue to be underpaid and exploited within these digital sweatshops. In fact, Kenya’s president, William Ruto actively encouraged Kenyans to seek remote work online after meeting with a college student who reported making Ksh.26,000 ($200) a week working on Remotasks. While the encouragement itself is benign, its power may not be. By encouraging Kenyans to seek remote gig work online, he inadvertently puts his stamp of approval on the nature of the work. Many Kenyans who would not otherwise have sought these opportunities will begin seeking them, leaving a larger percentage of the population vulnerable to tasks like explicit data annotation.
The allegations about Amazon’s Just Walk Out technology and the examination of data annotation and content moderation practices reveals a troubling pattern about the big tech AI business model. The AI global supply chain is efficient at best and neo-colonial at worst. It sees the Global South do the “dirty work” while the Global North reaps the benefits. This exploitation, driven by a hunger for profit, can only be mitigated by strengthening labor laws in the Global South. While the burden of adherence to ethical labor practices falls on the companies, it is the duty of each and every country to protect its citizens from harm.