We are entering a new era of technology dominated by Artificial Intelligence (AI). While years ago, AI was mostly in the hands of researchers, today we interact with AI on a daily basis, whether it’s through Google search results, Instagram recommendations, chatbot customer service, or the Siri digital assistant. Tech corporations are investing billions of dollars to scale up AI technology, and CEOs are making fantastical promises about developing artificial general intelligence (AGI) that surpasses human capabilities and that will address many of the world’s problems.
It is true that AI has incredible potential for positive impact, from assisting workers in grueling or rote tasks, to applications in medical imaging, which can improve early detection and diagnosis of disease. But the reality is that as tech companies rapaciously expand to maximize profits and market share, the AI boom has serious hidden costs for working people, our planet, and the communities we live in.
In this article, we provide an overview of the technology underlying AI, and analyze some of its hidden costs and their effects: from the impact on our environment from AI’s immense energy and water demands, to its impact on labor, especially exploitation in the Global South, to its inextricable links to militarism and imperialism.
Overview of AI technology
With the ease of access to AI tools like ChatGPT, many people are having a chance to interact with AI for the first time. From answering questions about math homework to getting help with writing an email, chatting with GPT can sometimes feel eerily similar to talking to a human. But what is behind a tool like ChatGPT?
AI models consist of large numbers of parameters, the values of which are optimized on the basis of many known examples of inputs and corresponding desired outputs. This statistical process is called “training” the model, and relies on machine learning techniques.
For instance, a language AI model like ChatGPT is optimized to predict the likely next word in a sentence as output, when given a partial sentence as input. Other AI models are trained to identify objects from image data. Powerful AI models have been developed for a multitude of applications including protein folding, image generation, and writing code. In each case, known examples of text, protein structures, images, or computer code were used to train the AI model. But the core scientific ideas behind these AI models have existed for many years. So what’s changed? Data and computation.
Recent advances in AI models and computing hardware have unlocked an exponential increase in the amount of data and computational power that these models are trained on. ChatGPT4 is trained on 10 trillion words (~ 3x larger than its predecessor), has 1.8 trillion parameters (~ 10x larger than its predecessor), and costs over USD 100 million to train (~ 20x times more than its predecessor). By scaling up these models, they are more accurate and have marketable applications that can be used by the public. But what does this expansion and scaling up mean concretely?
Energy
A first concern is the energy required to power the computers running these AI models. While most people interface with AI online through the cloud via our personal computers and devices, there are vast data centers filled with huge numbers of computers that are actually processing every AI query or request. In 2023, Google and Microsoft data centers each consumed more electricity than is consumed by over 100 countries. These data centers are filled with thousands of Graphical Processing Units (GPUs). GPUs are designed to run parallel calculations and are very efficient at performing the computations needed to run AI models. It’s no surprise that NVIDIA, the company that dominates the market for GPUs, is now one of the most highly valued companies in the world. And there is no sign of this energy consumption slowing down. NVIDIA has estimated it would sell 3.5 million of their newest GPU, which would consume electricity equivalent to 1 million U.S. households.
Google’s greenhouse gas emissions have increased by 48% since 2019. And electricity grids are having to turn to natural gas in order to meet the demands of data centers. The capitalist drive to maximize profits around AI generation has led to unfettered expansion of energy required for data centers and AI companies, which has exacerbated and set back decarbonization goals.
And this is already beginning to directly affect working people. Energy companies are raising electricity prices to fund building more energy infrastructure to support these data centers. Furthermore, these data centers have placed an incredible strain on the grid, increasing the chances of electricity blackouts during peak times. This risk is especially dire in California and Texas, where in the past we have seen how these blackouts have impacted vital infrastructure such as hospitals.
Water
But in addition to electricity consumption, these data centers also consume large amounts of water to cool the computers that drive AI. Researchers at the University of California, Riverside, estimated last year that global AI demand could suck up 1.1 trillion to 1.7 trillion gallons of freshwater by 2027. These data centers, as well as manufacturing facilities for the semiconductor chips that are necessary for AI computing, are often geographically concentrated in areas where water isn’t plentiful. For example, many new data centers are being built in California, which faces regular droughts, and in Arizona, where the Colorado River is drying up.
And this data center expansion is not just in the U.S. Google is looking to expand internationally as well, including building data centers in Uruguay. Uruguay is already facing its worst drought in 74 years. There have been widespread protests against additional data centers which would further burden their water supply and risk people’s access to fresh drinking water. One of the slogans of their movement opposing these expansions is “this is not drought, it’s pillage,” which you can find scrawled on walls across Montevideo.
Corporations are also very aware of the detrimental impact that water consumption is having on communities, so it is no surprise they have not been transparent with their consumption. In one instance, in Oregon where Google runs three data centers and plans two more, Google filed a lawsuit, aided by the city government, to keep their water consumption a secret from farmers, environmentalists, and Native American tribes. After they faced pressure to release the data, they caved and the records were made public. They showed that Google’s three data centers use more than a quarter of the city’s water supply.
Militarism
While the tech industry likes to highlight the positive, or at least the innocuous, impacts of AI, these companies are also simultaneously developing the dangerous and violent applications of AI. The Pentagon has recently requested USD 1.8 billion for the fiscal year to support the delivery and adoption of AI-enabled capabilities. And this money is often flowing to companies like Google, Microsoft, and Palantir, and universities like MIT and Stanford in the form of defense contracts. Competition over this technology, particularly around control of semiconductor chips that underpin AI, is also driving a new front in U.S. aggression towards China.
We already see the use of AI in warfare. Project Maven is a military program being run by the U.S., which uses AI to distinguish people and objects on the battlefield. This program was previously in collaboration with Google, but after they faced pressure from employees, they dropped the program and now the U.S. military works with Palantir. Gospel and Lavender are two AI systems used by the Israeli Occupation Forces. “Gospel” is a system for marking buildings to bomb that it says Hamas militants are using. And “Lavender,” which is trained on data about known militants, then parses through surveillance data about almost everyone in Gaza–from photos to phone contacts–to rate each person’s likelihood of being a militant. While the IOF claims that this is still gated on a human making the final call, Israeli soldiers told +972 that they essentially treated the AI’s output “as if it were a human decision,” sometimes only devoting “20 seconds” to looking over a target before bombing, and that the army leadership encouraged them to automatically approve Lavender’s kill lists a couple weeks into the war.”
Labor exploitation
While the mainstream narrative around AI characterizes AI models as systems that can operate entirely autonomously and are freeing workers by building machines that can do the boring and repetitive tasks, this is actually quite far from the truth. Instead these AI companies are building their AI models by treating many workers like machines. Training these models requires enormous amounts of data, and much of that data is cleaned and annotated by humans. Tech companies leverage the economic disparities between regions and this work is often outsourced to workers in the Global South including Syria, Argentina and Kenya, where workers are paid less than USD 1.50 per hour, with little job security and no clear path to upward mobility, and no protections for workers rights.
And that’s not even considering that the work itself is extremely repetitive and these workers are under careful surveillance and are punished if they deviate from their prescribed repetitive tasks. Some of the types of tasks might involve watching content and assessing whether it needs to be flagged. This means the workers must watch video sometimes containing suicide, murder, child abuse, and sexual assault and many workers report having developing stress and anxiety disorders from being constantly exposed to this content.
Of course, this form of labor exploitation is in addition to the likelihood that corporations will use AI to deskill or eliminate huge numbers of jobs globally in search of greater profits.
Conclusion
AI is already transforming the lives of working people across the world and how society functions. However, in the hands of corporations whose interests are in profit-making and competition, the hidden costs of this technological advancement will continue to be paid by working people across the globe. Only by centering the interests and wellbeing of the majorities, can the implementation/use/adoption of AI take place without sacrificing our planet, our communities, and our livelihoods in the process.
Nishad Gothoskar is a Ph.D. candidate in the Computer Science and Artificial Intelligence Laboratory at the Massachusetts Institute of Technology who specializes in probabilistic methods in computer vision and robotics