The age of social media has made it easier than ever to spread misconceptions and misdirection. Bad arguments seemingly go through seasons like the flu. One idea currently in-season is that artificial intelligence is uniquely damaging to the environment. Accusations abound against AI’s impact on our environment, emanating both from popular consciousness and academic institutions.
These claims vastly overstate the negative impacts of AI. Below are three popular arguments and how they misunderstand the economics behind our relationship to the environment.
Argument 1: AI Uses Excessive Resources
The most common concern around AI is its usage of resources—mainly electricity and water. Detractors claim that AI’s energy usage is dangerous, excessive, and unnecessary. The only way to train and operate LLM models like ChatGPT, Claude, Gemini, etc. at scale is through large data centers. The two resources these data centers need to operate are electricity and water. Because of the constantly-increasing scale of AI usage, their consumption of these resources is likely only to rise in coming years. At the very least, there should be serious concern about this environmental profligacy.
Is there any truth to these claims? First, it is true that LLM models do use considerable amounts of electricity and water, but the actual usage is much less than commonly believed. Many of the popular figures of AI’s electricity and water usage are based on older estimates which are now obsolete. As confirmed by Sam Altman of OpenAI, Google’s internal research, and independent analysis, a single LLM query uses ~.3 watt hours of electricity. This is 10x less than previous estimates, which pegged a single query at consuming 3 watt hours. Water usage is significantly lower as well, which Sam Altman reports as ~.26 ml of water per query.
How much electricity and water do these programs use in the aggregate? Unfortunately, most companies haven’t made this data public, but we can use the data we have to get rough estimates. OpenAI reported to Axios that ChatGPT processes 2.5 billion queries per day. Using our per-query estimates, we get the following numbers: 750 million watt hours per day; 650,000 liters of water per day.
When viewed in isolation, these numbers look much scarier than they really are. Let’s compare the electricity usage of LLMs to other daily tasks:
- One day of running your refrigerator = 1,600-3,300 ChatGPT queries
- Leaving your oven on for one hour = 7,600 ChatGPT queries
- Drying a light load of laundry = 8,300 ChatGPT queries
The comparison in terms of water is even more damning. The US Geological Survey reports that the US uses over 1.2 trillion liters of water every day. This means that ChatGPT accounts for about .00000043 percent of the US daily water consumption.
These aggregate numbers only represent ChatGPT, but it’s unlikely that any competing LLMs are using significantly more electricity or water—especially concerning ChatGPT’s position as the foremost AI program.
If AI’s usage of electricity and water are grounds for concern, we have a long list of much more pressing concerns.
Argument 2: AI Uses Scarce Resources
Even if the electricity and water burden of AI isn’t inordinate, it still uses other resources that cannot be replaced. The GPU chips that LLMs require are built with a variety of rare materials that we don’t have an infinite supply of. These include gold, tungsten, copper, aluminum, etc. Given the dubious value of AI at its present state, shouldn’t we consider the possibility of limiting its production to conserve our natural resources?
First, we are in no danger of the Earth “running out” of precious or rare metals any time soon. For instance, the USGS estimates that 54,000-64,000 metric tons of gold lie in proven underground reserves, in comparison to roughly 210,000 tonnes of gold that have been mined throughout history. Similarly, the USGS estimates that 6.3 billion tonnes of Copper remained with only 700 million tonnes mined so far.
But won’t we eventually run out of these resources? No matter how well we ration them, their supply is still finite. Shouldn’t that concern us?
Finite resources are always a concern, but history shows us that markets are adept at dealing with resource constraints and shortages through two methods: price-related rationing and incentivizing substitutes.
As the supply of a resource decreases, the price will be pushed upwards over time. As with any price increase, this results in less of that resource being used. Additionally, the resources that are used are directed to the highest-valued outputs. Simultaneously, this higher price creates a profit opportunity for anyone who can bring a substitute to market. If a substitute is discovered, they can undercut the higher price of the resource in shortage, creating an instant market for their product.
An example from history shows this market mechanism at work: in the 16th-17th century, wood was by far the primary source of fuel for heat and cooking. Their usage of firewood was so great that trees were increasingly difficult to find and source. The price of wood was ten times more expensive in 1620 than in 1540. As trees became more sparse, Englanders were desperate for a cheaper alternative. At the time, coal was already known as a potential fuel source, but it wasn’t widely used. But when the price of wood skyrocketed, it was an increasingly-tempting alternative. With additional investment and innovation, coal became cheaper, more widely available, and eventually replaced wood as the burning fuel of choice. Today, England is once again covered with trees and forests.
It is true that Earth’s precious metals are scarce, but so are all other resources. Markets and market-driven innovation have dealt with resource shortages before, and there’s no reason to think they couldn’t step up to the challenge again.
Argument 3: AI Has a Devastating Environmental Impact
Even if the impact of AI is minimal and there is little danger of running out of natural resources, the operation of AI still causes environmental damage. Because of the resources it relies on, it’s complicit in acts like mining, deforestation, pollution, etc. that are choking our planet. These ecological damages will only scale as AI usage increases. Regardless of resource use, AI’s deleterious impact on the environment alone is a basis for serious public concern.
This perspective is a perfect example of what Alex Epstein calls the “anti-impact framework”—a worldview that elevates preservation of the environment above all else. Minimizing our impact on the environment is elevated to the highest end. The conclusion of this elevation is that we should abstain from anything that changes or alters the natural environment.
Despite its immense popularity today, this framework is profoundly anti-human. It is only because we changed and manipulated our environment that our modern achievements of increasing calories consumed, worldwide GDP per capita, and life expectancy are possible. While I have nothing against the environment, I suggest an alternative viewpoint: we should elevate human flourishing as our highest good.
With this as our operating principle, the environment isn’t something for us to revere and leave alone, but mold according to human needs. If that requires cutting down trees, digging for minerals, or genetically modifying crops, we should do it. The environment isn’t an all-loving provider that gets us whatever we need, and we aren’t parasites leaching off its bounty. The environment is just where we live, and we should change it according to our own ends.
Of course, this doesn’t mean we never preserve our environment. Parks, nature reserves, natural landmarks, etc. are all good things, but they are good because they serve human interests and wants, not because they are sanctuaries to “Mother Nature.”
Conclusion
As with all temporary fashions, the flawed arguments against AI’s environmental impact will eventually fall out of favor. Even so, they serve a valuable purpose: exposing economic ignorance around energy. While the ultimate impact of AI on our world is yet to be seen, there is one thing we can know for sure: its environmental impact won’t be the reason it fails.
