HaoTech
KEY INSIGHTS
AI companies are taking advantage of affordable labor, from impoverished workers in underdeveloped nations to interns and even prison inmates, to feed their algorithms with enormous data at meager wages. What’s the hidden aspect of data labeling, and why is human involvement so vital in training AI?
Artificial intelligence requires ingestible datasets, and a significant amount of manual human intervention is highly beneficial, if not outright necessary.
However, data labeling is a monotonous and often grueling task, and several companies are exploiting low-wage labor to supply their AIs with enormous quantities of data. From impoverished workers in underdeveloped nations to interns and even prison inmates, companies proclaiming their commitment to “ethical AI” might be employing vastly underpaid workers in a less than ethical manner.
Do humans play a crucial role in training AI? Who are the most underprivileged data labelers and annotators laboring in digital sweatshops, and where do they originate? Are we certain there are no other alternatives to human intervention when feeding AI with training data? Let’s delve into these questions.
The Human Element: An Indispensable Ingredient in AI Training
The ‘human touch’, often referred to as ‘human-in-the-loop’ models by experts, is crucial for maintaining data quality during AI training.
In 2018, a fatal accident involving a self-driving car and a woman walking her bicycle across the street highlighted the limitations of machine learning-based models. The AI could recognize a pedestrian or a bicycle as separate entities but faltered when they appeared combined – an unexpected scenario it hadn’t been trained for. These models are inherently brittle and lack the flexibility of human decision-making when confronted with unfamiliar situations.
Humans are required to address these ‘edge cases’, where informed decisions are crucial, safeguarding AI from its inherent brittleness that could lead to malfunction when faced with the unknown. But human involvement is necessary for more than just this reason.
When companies resort to datasets that do not require human labelers, such as machine-generated or structured data, the results have been less than satisfactory.
Models trained exclusively on outputs generated by other AIs tend to behave erratically over time. A study from Rice and Stanford University noted a phenomenon called Model Autophagy Disorder (MAD) where the quality of outputs, like images or videos, starts to degrade, becoming increasingly strange and absurd. This phenomenon is likened to a neurodegenerative disorder affecting the ‘brains’ of machines, drawing parallels with the infamous ‘mad cow disease’ that affected self-consuming cows.
**The Unseen Labor Force: Workers from Low- and Middle-Income Countries**
One of the most glaring issues of globalization is arguably the outsourcing of jobs to countries where wages are extremely low and working conditions are exploitative. The simpler and less specialized the tasks required for a job are, and the more these tasks can be performed remotely, the easier it is for large organizations to exploit outsourcing. Workplace overcrowding and high turnover rates are irrelevant if anyone can perform the task without any substantial training.
Predictably, this new workforce of 21st-century laborers comes from the less developed economies of Africa and Asia, where wages are low and workers’ rights are often minimal. A recent investigation by TIME revealed that OpenAI, the creator of the globally renowned ChatGPT, outsourced workers from Kenya, Uganda, and India to cleanse its chatbot of toxicity, violent language, and bias. In addition to dealing with disturbing datasets (more on this later), the Kenyan workers received a meager wage of between $1.32 and $2 per hour for their contribution to this multi-billion dollar market—while the agency managing the work was allegedly paid $12.50 an hour per worker.
‘Clickworkers’: The Global Labor Force
African workers aren’t the only ‘clickworkers’, as they’re often referred to, who are hired for astonishingly low pay. In the Philippines, thousands of young, unskilled workers spend their days distinguishing lampposts from pedestrians in videos used to train self-driving cars, identifying celebrity photos, and editing text snippets. Their pay? A mere $6 to $10 a day. Do they have basic workers’ rights? Certainly not, as they’re hired through freelancing platforms that outsource their work to larger AI companies.
These same platforms often withhold their payments for a week, seize them for any alleged violation without any recourse, or ban them if they attempt to log in from a different device. Because your wage is determined by your country of origin, using any VPN can instantly cost you your job – and if your pay is already close to the poverty line, sudden unemployment can be devastating.
21st Century Forced Labor: The Lowest of Lows
How does it look in more affluent countries? Not much better. Different locations, different strategies, but the end result remains the same: human exploitation. In China, parts of
the AI industry have struck an unsavory deal with vocational schools. Students are required to perform tedious, patience-testing data labeling and annotation tasks in order to graduate. They are lured into internships advertised as “career-enhancing jobs” that are nothing more than low-paying, repetitive assembly-line tasks. After the greedy vocational schools take their cut, the pay is barely enough for a meal, falling short of even the local minimum wage.
In the advanced Western world, we may have reached a new low in digital servitude. In Nordic countries, where data must be collected in local languages spoken by a small population, such as Finnish or Danish, it’s more challenging to employ underpaid African or Indian workers. So, who better to do the job at a fraction of the cost than prison inmates? Earning a pitiful €1.54 ($1.65) per hour in a country where a Starbucks Espresso costs €2.8 ($2.99) is well below the poverty line. But then again, if you’re a prisoner, what could be better than a job that will “prepare you for the digital world of work” once you’re released, as the prison system proclaims?
The Unseen Horrors of the Clickworker’s Job
Whether in the East or West, North or South, the job of a clickworker is far more tedious and punishing than one might imagine. Sitting in front of a computer and clicking away to earn money might seem better than manual labor under the sun, but that’s not always the case. The work carried out by data collectors, labelers, and, most distressingly, social media moderators can be quite horrifying.
As previously mentioned, the Kenyan workers employed by ChatGPT were tasked with removing toxic content from the chatbot. To do so, they first had to identify the toxicity, which meant reading, watching, and experiencing it. This toxicity emanates from the darkest corners of the internet and, often, the darkest depths of the human mind. Social media moderators tasked with feeding content moderation algorithms are frequently exposed to terrifying images, videos, and content filled with violence directed at humans and animals, pornographic content, gore, and heart-wrenching abuse. All this for a meager wage and in working conditions that often border on abysmal.
The Somber Reality: A System Designed for Exploitation
Both small and large tech companies employ various exploitative tactics to ensure that employee rights are not guaranteed, a fact that is well-known. However, the system itself is not designed to improve workers’ conditions over time. The job goes to the lowest bidder, and since outsourcing agencies and freelancer platforms take a share of the workers’ pay, their final wage is even lower. Middlemen ensure there’s a steady stream of people ready to take on the job, who can be hired or fired at any time, and who live under the constant threat of a ban that could prevent them from securing future work. Labeling work offers no path for career advancement: there’s no real skill to be learned, no way to market oneself as a professional, and no opportunity to earn a higher wage over time.
We often envision a world controlled by machines as a dystopian reality where our synthetic overlords force us into gray, monotonous routines. We are merely cogs in a massive machine that consumes us, nameless, faceless, and devoid of identity. Yet, we seldom pause to consider that today’s world already mirrors this grim scenario for a vast majority of less fortunate individuals across the globe. It begs the question: how could machines possibly make things worse than they already are?
