It is not difficult for any high school or college student these days to rattle off a dozen reasons that they have been told to avoid using ChatGPT or artificial intelligence (AI) for their schoolwork. These reasons are not arbitrary: concerns about inaccurate facts and a desire for students to go through the traditional process of completing assignments are certainly fair. However, the problem with focusing the AI discourse so narrowly on education is that it causes us to think of and treat AI as a concept rather than a tangible thing with real-world implications, and to only consider the ethical implications of AI in a very specific context.
If we condemn only certain uses of AI — to write graded assignments for you — we neglect to think about the impacts of having it, say, generate a title for a midterm paper or write an email to your professor. It’s no secret that computation is energetically expensive and dependent on water for cooling purposes, but conversations surrounding the huge environmental impact of AI specifically have just recently come into the mainstream.
The excessive water consumption of companies that create AI is undeniable. For example, Microsoft’s water consumption increased 34% from 2021 to 2022, when they used a reported 1.7 billion gallons: a spike related to their development of AI technology. I should note here that widely spread claims that ChatGPT uses roughly sixteen ounces (473 mL) of water for every 100-word email it is prompted to write are hard to verify and are imprecise. For example, different data centers require vastly different quantities of water to perform this task, with numbers ranging from 235 mL (in Texas) to 1468 mL (in Washington). Nonetheless, the contentious nature of this claim does not negate the issue of AI’s water consumption.
In a world where more than one in four human beings do not have consistent access to clean water and sanitation, what should be prioritized? Should rapid technological advancement which makes the wealthy wealthier and brings unnecessary convenience to citizens of the most privileged countries in the world come before ensuring that everybody on this planet has sufficient access to basic necessities?
The inequity runs deeper than just water, though. Ultimately, the adverse climatic effects of AI will not first hurt the individuals who actually use the technology. Instead, AI will do the most immediate and intense harm to the populations who did not have water to drink in the first place, while data centers are guzzling it up. These vulnerable populations will be doubly victimized by technology to which they did not even have access.
The broader effects of AI go beyond water and energy usage in more ways than one. Part of the significant risk of AI is that we do not know what advancements will be made and when, so it is difficult to take preventative measures against potential adverse effects the advancements will cause. Protective legislation cannot predict how AI will develop, and it risks becoming obsolete soon after being passed.
AI’s climate-change-related dangers are pervasive. One possibility is that AI will soon be used to make improvements to the development of self-driving cars, which will cause more people to prefer driving as opposed to other transportation options in the long term. If this potential trend were to continue for a long enough time on a large enough scale, greenhouse gas emissions and/or electricity usage (by electric cars) would increase greatly. Another of many possibilities is that AI-produced misinformation about climate change will lead the public to be less concerned about the matter than they might be otherwise. This possibility is substantiated in part by the already huge issue of AI creating increasingly believable and massive amounts of misinformation.
It is hard to pinpoint the total amount of damage AI does, in part because many major search engines and even social media platforms have some form of an AI chatbot nowadays. It is difficult to compile the impact of all AI development and use across such a multitude of corporations. However, we do know that whatever the electrical cost might be, it is vast.
The amount of energy consumed by any given chatbot is partially static and partially dynamic. The static portion of the quantity is consumed in the training of the bot, which is a one-time process for each bot that is created. ChatGPT-3, for example, consumed nearly 1,300 MWh (megawatt-hours) of electricity throughout its training. To put that measurement in perspective, the average American home consumes just ten MWh of electricity on an annual basis. Additionally, it’s important to remember that over half of the electricity consumed in the United States still comes from fossil fuels; in the training of ChatGPT-3, over 500 tonnes of CO2 were released into the atmosphere. The dynamic portion is the amount of energy that will be necessitated by an unpredictable number of queries, which will be discussed further below.
While official figures about what it took to train ChatGPT-4o have not been released, based on the time it took to train, estimates hover anywhere from 1750 to 7200 MWh. This range is so wide because it is difficult to be certain exactly how many parameters the model has, which is directly related to the time and power it took to train.
The number of parameters a model has is important for three main reasons. Firstly, as mentioned above, a model with a larger number of parameters will require a longer and more energetically expensive training period. Secondly, the number of parameters a model has is directly proportional to its level of sophistication. Thirdly, the more sophisticated a model is, the more energy will be required to carry out each individual query entered by a user.
Advancements in AI greatly increase the dynamic portion of the energy consumed by AI tools. People are more likely to use more sophisticated products in the first place and are presumably more likely to use a sophisticated model more frequently than they might use a less advanced model. Therefore, with increased levels of sophistication, the energy required per query will be larger (for example, ChatGPT-4 consumes 0.0005 kWh (kilowatt-hours) per response, as opposed to ChatGPT-3’s 0.0003 kWh), the user base will be greater and the average number of queries entered by each user will presumably be larger. It seems that we as a species are digging ourselves into a deeper and deeper hole with the advancement of AI.
I myself am no Greta Thunberg, nor do I believe that the deeply important responsibility to protect the environment should fall solely to us as individuals. I do believe, though, that in a world where governments and corporations refuse to act as stewards of the earth in any meaningful way, we have an obligation to do what we can. So, the next time you’re too tired of looking at that Google Doc to think of a title for your essay, consider asking a friend for some help rather than a bot. Each and every one of us should make a reasonable effort to take steps towards a better future — or, more attainably, one that isn’t as bad as it might have been otherwise.
