Blood, Sweat, and Tears Also Flow Through the Veins of AI [AI Mistake Notebook]

Is Artificial Intelligence Truly "Artificial"?
Exploring the Human and Environmental Costs of AI

Editor's NoteExamining failures is the shortcut to success. 'AI Mistake Notebook' explores cases of failure related to AI products, services, companies, and individuals.

The flowing stream water being scooped up by the robot's both hands. Photo by Gettyimagesbank

What kind of image comes to mind when you hear the word artificial intelligence (AI)?

Most people probably imagine the cold, dry texture of metal or the precise, linear movements of a robot. The word "artificial" gives a sense of something in contrast to the natural. On the other hand, the word "human" evokes warmth, emotion, and a sense of softness.

This contrast is a common impression. People tend to think of artificial intelligence as belonging to the rigid world of computer code and cold algorithms, while humans exist in the realm of warm emotions and intuition. But is this really the case? Is artificial intelligence truly only "artificial"?

AI services such as ChatGPT and Claude are learning from all digitized texts, images, and videos, absorbing human knowledge, experience, and even emotional nuances. However, this learning process is not purely mechanical.

Exploiting the Earth: Artificial Intelligence and Natural Resource Consumption

In January 1992, Deng Xiaoping, who was leading China's reform and opening-up, visited the Baotou region in the Inner Mongolia Autonomous Region and said, "If the Middle East has oil, China has rare earths." The photo shows a rare earth mine in Inner Mongolia, China. Photo by Reuters Yonhap News

Electricity flows through the veins of AI. The hardware and software that make up AI operate and compute using electric power. These AI systems literally consume an enormous amount of electricity.

According to the International Energy Agency (IEA), global data center electricity consumption was 460 TWh in 2022 and is expected to more than double to 1,050 TWh by 2026. Samil PwC Management Research Institute even published a report titled "AI Grows by Consuming Electricity."

Data centers are often associated with the cloud. The cloud is a service that allows users to rent computing resources via the internet as needed. The cloud operates through data centers. The word "cloud" conjures up images of cleanliness and eco-friendliness. However, this is not always the case.

Data centers are among the world's largest electricity consumers. Because data centers generate heat, they also require water for cooling. The enormous water consumption for cooling can even lead to local water shortages. In China, 73% of the energy for the data center industry came from coal, and in 2018 alone, it emitted 99 million tons of carbon dioxide.

Rare earth minerals are essential for manufacturing the hardware (such as CPUs and GPUs) needed to run AI systems. The problem is that mining, refining, and processing rare earths is extremely complex and causes severe environmental pollution.

According to the United States Geological Survey (USGS), China accounts for more than 70% of the global supply of rare earths. Is this because rare earths exist only in China? Or is it because only China has the technology to mine them?

To make mined rare earths usable, they must go through hazardous processes involving large amounts of sulfuric and nitric acids. This endangers not only the environment but also the lives of human workers. The reason China can monopolize the supply of rare earths is that it is the only country willing to bear such costs. In other words, rare earth mining is competitive where labor is cheap and environmental regulations are lax.

Labor Exploitation: Automation Fueled by Human Beings

The large-scale image database ImageNet. The task of defining and classifying what each of these images contains must be done by humans. Today, ImageNet has made a significant contribution to the advancement of deep learning models. ImageNet

The extent to which AI consumes the Earth's resources and negatively impacts the environment is relatively well reported and widely known.

However, in addition to natural resources, maintaining and developing AI systems requires countless low-wage workers. "Labor exploitation for AI" has received less attention compared to environmental exploitation.

AI must learn from data, and the value of AI is determined by the quality of the data it learns from. One of the most critical datasets for AI learning is "ImageNet."

For AI to recognize an image of an apple as an apple, it must first learn from countless images of apples. This means that training AI on images requires thousands or even billions of photos, and there must be preliminary work to classify and label these photos. ImageNet is the image database that made this possible.

Behind this work were countless low-wage workers. IT workers in developing countries, receiving extremely low pay, labeled each photo with its title and identified what was depicted in each image.

Modern-Day Slavery Underpinning AI

An image depicting numerous low-wage workers performing image classification tasks in front of monitors, as described by AI. DALL-E3

The ImageNet project was carried out through "Amazon Mechanical Turk." Mechanical Turk is a crowdsourcing platform, which, simply put, is a service for buying and selling low-wage labor. It connects those who need data classification, editing, and labeling tasks with people willing to provide such labor.

Nearly 50,000 workers from 167 countries were mobilized for the ImageNet project. They organized and classified about one billion images. What we consider AI's intelligence is, in fact, the result of the repetitive labor of many people.

It is easy to think that classifying and labeling images and data is a simple task. However, these workers are not just sorting photos of cats and dogs. They must directly view all sorts of data, including violence, torture, child abuse, disaster scenes, animal corpses, blood, and filth. Exposed to such information for long hours every day, workers suffered psychological trauma so severe that normal daily life became impossible.

Data workers in Kenya had to review brutal and horrific images and texts for more than eight hours a day, yet their wages were only about 1,600 to 2,400 won per day.

Lilly Irani, a professor at the University of California, describes digital technology and AI systems as "automation fueled by human beings."

In an open letter to U.S. President Joe Biden in 2023, Kenyan AI data workers claimed that "U.S. big tech companies are systematically exploiting African workers and ignoring local labor laws." They described their working conditions as "modern-day slavery."

"We must watch murder, decapitation, child abuse, rape, pornography, and bestiality for more than eight hours a day. We do this work for less than two dollars an hour."

For the Earth and Humanity: Sustainable AI

An image depicting development that considers the Earth and the environment. Getty Images Bank

The saying "electricity flows through the veins of AI" is true. It is also true that the blood, sweat, and tears of humans flow through AI's veins.

The AI infrastructure we use every day is deeply infused with human labor and the Earth's resources.

We must continue to advance AI while making every effort to minimize its environmental impact. Data center operations should utilize renewable energy as much as possible, and prioritize infrastructure with high energy efficiency.

Workers involved in AI development must receive fair compensation and be provided with safe working environments. Mary Gray and Siddharth Suri, former Microsoft researchers, emphasize in their book "Ghost Work" that "hidden labor must be brought out of the shadows." Psychological support is also necessary for those handling content that can cause trauma.

Artificial intelligence can no longer hide behind the word "artificial." AI is not the exclusive domain of a few big tech companies; it is the result of the labor of people and the resources of the Earth.

If we accept AI as an unavoidable reality, we must also face these uncomfortable truths. Only then can we create a fairer and more sustainable AI.