Artificial Intelligence (AI) is a hot topic, and honestly, I am tired of people discussing it. I’m tired of AI in my google and Instagram searches. I am tired of Character AI. I am tired of Deepfakes. I am tired of accidentally pressing a button on my keyboard that pulls up the most hideous chatbot that is Microsoft Copilot.
If I could go my whole life without hearing the term “Generative Artificial Intelligence” ever again, I could die happy.
Of course, the common ethical gripes most AI critics highlight are how it steals from artists and sucks the soul out of human creation, enables CEOs to fire mass amounts of people, and panics college professors and high school teachers into trying their best to curb the rampant cheating within academia.
However, I am interested in the massive negative environmental impact of AI.
According to a study done by Alex De Vries in Joule, a monthly peer-reviewed journal, ChatGPT-3’s training phase required 1,287 Megawatt hours which is more energy than what Japan, Mexico, and Saudi Arabia used in 2023 combined.
ChatGPT-3 is the third edition to OpenAI’s Language Learning Model (LLM). A single ChatGPT query and Google AI search both use 3-watt hours of energy per response, which is ten times more than what a Google search would use without AI.
In 2023, there was an average of eight billion Google searches per day, or 98,379 per second, according to data from Statistica. Which means Google AI search results could use as much as 295,138 watt-hours per second, which is enough to power the average Massachusetts house for 15 days.
So, we know that AI uses a significant amount of energy, but what does that mean for the environment? As stated by Towards Data Science, ChatGPT has a daily carbon emission of 414 kilograms of carbon dioxide or equivalent. That is enough to drive 1,059 miles in an average car. Yearly, that would be 151,110 kilograms of CO², or enough to power 19 average American homes.
All this energy generates massive amounts of heat, meaning they need some way to cool it down; data centers use cold water and outside air assistance to do so.
To cool data centers down, AI needs two types of water usage: water consumption and water withdrawal. Water consumption means that once the water is used, it evaporates into the environment or becomes unusable. Withdrawal means the amount of water that is withdrawn from a freshwater source.
Since the training phase for a generative AI program uses an extreme amount of energy, the phase also needs an extreme amount of water to cool the data center down. According to Professor Shaolei Ren at University of California, the training phase for a large language learning model like ChatGPT “can consume millions of liters of fresh water”. To put that in perspective, one million liters of fresh water can fill 400 Olympic size swimming pools.
Shaolei Ren has also estimated that an average conversation with ChatGPT-3 or around 10-50 responses can consume up to a half liter or a standard plastic water bottle’s worth of freshwater.
Google’s yearly report has said that their water usage has risen by 20 percent from 2021 to 2022 and that 15 percent came from areas with “high water stress”. While Microsoft’s report said water usage has risen by 34 percent in the same period. It also said that 42 percent of the water came from places with “high water scarcity”.
However, scholars and activists have been criticizing the lack of transparency that companies such as Google and OpenAI have on exactly how much water and energy their data centers consume. Most of the current information comes from estimations and comparisons to similar technologies.
So, before you ask ChatGPT a question about your homework, to write an essay for you, or a question you could simply Google, I want you to think about the impact that search could have on the environment.
Because currently, no one knows for certain the exact impact that those searches could have. I do not think using an unknown amount of electricity and water is worth it to know a robot’s opinion on that paper you have to write.
