The Environmental Cost of AI and How the World Can Mitigate It
It’s no secret companies like Google, OpenAI, and Microsoft are dominating the web with chatbots like ChatGPT and Bing AI, and Google Bard. Competition has led the companies to try to develop the best AI in an “AI race” like the Space Race of the 1970s and 80s.
The recent advances in AI have shocked the world and some say that the end of days is near. Movies like Terminator have portrayed AI as a military threat that will try to conquer the world and wipe out or enslave humanity. So far, AI will not be in the military due to its high vulnerability and potential to be hacked. However, AI currently poses a threat to something that no one predicted: The Earth’s climate.
Understanding the Threat
Understanding the threat starts with understanding the cause. AI works like a child: thousands of sample questions and answers are fed to it so that it can understand and learn from them. The AI then sets its own parameters to evaluate and respond to questions which it “learned”. This “knowledge” is known as weights, which e store the AI’s analysis of the sample data and affect how it interprets and responds to input. It’s like the consciousness of the AI, with more samples leading to a more intelligent consciousness. For example, GPT-4 by OpenAI is trained on about 100 trillion samples. If GPT-4’s weights were used on a language model which was being developed, it would already have the “knowledge” required to generate responses to any text-based or image-based questions fed to it without needing any training. However, AI models are often trained independently by researchers and corporations to fit their specific needs. In addition to this, AI models often also must be trained and tested multiple times to fine-tune their parameters so the models can set better weights and reach a desirable accuracy of output. For example, weights from an AI designed to generate images would not be transferrable to a text-based AI like GPT, or the weights for a text-based AI chatbot designed to be a medical assistant would not be a good choice for a chatbot which answers physics questions.
This training, however, requires extreme amounts of money, time, resources, and energy, such as buildings as offices for the people who type in the questions and answers and their salary. Many large corporations have their own AI divisions with several people who manually train them, which undoubtedly does create several jobs for people and therefore boosts the economy. However, training these AI models also greatly increases the amount of carbon emissions by making companies race against each other to get the most workers to enter the most parameters into training their AI to make the “best AI”. The use of AI is almost as harmful if not just as harmful as its training. According to MIT’s Tech Review Journal, “Generating one image [using Dall-E] takes as much energy as fully charging your smartphone.”
Potential Solutions
So, how can AI be utilized in a sustainable way which preserves the planet? Well, the first and easiest to enact solution is to share AI algorithms to find more efficient ones. Stanford University’s Human-Centered-AI Journal, for example, says that “more than 200 students in a class on reinforcement learning were asked to implement common algorithms for a homework assignment. Though two of the algorithms performed equally well, one used far more power.” If a single student at Stanford was able to make an algorithm more efficient than another, corporations almost absolutely would have more efficient algorithms than each other at their disposal. If companies like OpenAI, Google, and Microsoft shared their algorithms, it would kill two birds with one stone by allowing them to pick the most efficient algorithms to reduce power consumption while also drastically reducing emissions from training their own AIs and saving the planet.
Another government-centered solution is imposing a “Training Tax” upon artificial intelligence researchers and corporations, similar to Sweden’s carbon taxes, which have decreased the nation’s carbon output by a whopping 27 percent. The idea would be to reduce the amount of training by making a system which has researchers pay a “training tax”. This encourages researchers to think before they train their systems because of the cost associated with training it, similar to how children often hesitate to play a game that they don’t know at an arcade and lose some money in the process. This would encourage less training and therefore reduce the greenhouse impact of AI profoundly.
AI research is a developing field, and while it is still developing, a line of research in making AI more energy-efficient must be introduced. If more efficient computers were never developed, then computers would still run very slowly and take countless hours and lots of energy to do basic things. If more efficient vehicles were never developed, modern cars would still require either coal or gas tanks in their trunks and need to cool down every few hours. Because of the abundant historical evidence that efficiency is the best way to further develop a technology, it is about time that Artificial intelligence also introduces branch in its line of research dedicated exclusively to making algorithms more efficient.
Data centers are a large part of AI and modern networks in today’s world. Because of the increasing dependence of AI on data and lucrative emerging fields like Data Analysis and Data Science, they are currently being used more than ever to the point where databases are now “accounting for around 1.8% of electricity use in the United States.”²
As the graph presented above shows, these databases not only use about 2 percent of the US’s electricity, but also uses millions of gallons of water for cooling and hydroelectric purposes. Innovations in cooling and server optimization should be prioritized due to the large amount of usage of databases already and the increasing prominence of data in today’s world because of the tremendous number of potential reductions in carbon emissions.
Even though AI has an impact upon the environment every time it is used, it can still be a force for good. Energy grids are often optimized based on previous trends, and AI is a master at learning from trends and suggesting changes to things based on them. Because of this, AI is often used to help make decisions in the power grid. AI also recognizes more complex patterns, such as weather with Google’s GraphCast model for more accurate and faster weather forecasting. Because of its ability to recognize patterns, AI is miraculously able to reduce its own carbon footprint through a process called carbon-aware computing, which automatically changes computing tasks to other parts based on the availability of renewable energy sources. This could be extremely helpful to reduce AI’s carbon footprint because AI is on the track to becoming exponentially more advanced and has the potential to save massive amounts of energy as it matures.
The impact of these initiatives will only be seen, however, if AI companies are transparent about their carbon footprint. So far, Microsoft is the only company with an AI branch which has made a statement that it will commit to lowering its carbon footprint intentionally. Other AI-producing companies are only barely passing the government regulations, which is why either they must be made stricter, or AI companies must be morally responsible enough to clean up the mess they made.
In conclusion, the AI revolution has the potential to provide positive benefits, but also can be very harmful to the environment if improperly managed. It’s important for corporations to work together instead of competing against each other and focus on developing a swift, carbon-efficient AI. Policymakers will also play an important role in regulating the field with the introduction of new standards and taxes upon the new field. If the AI community fails to regulate itself, the short-term gains of AI will come with tremendous long-term repercussions.
References
1- Mehta, J., Jarenwattananon, P., & Shapiro, A. (2023, June 26). Behind the secretive work of the many, many humans helping to train AI. NPR. https://www.npr.org/2023/06/26/1184392406/behind-the-secretive-work-of-the-many-many-humans-helping-to-train-ai
2- The Algorithm. (2023, November 10). Generating one image takes as much energy as fully charging your smartphone. MIT Technology Review. https://www.technologyreview.com/2023/11/10/1036149/ai-carbon-footprint-gnome-deepmind-materials-discovery/
3- Jonsson, S., Ydstedt, A., & Asen, E. (2023, November 22). Looking back on 30 years of carbon taxes in Sweden. Tax Foundation. https://taxfoundation.org/research/all/eu/sweden-carbon-tax-revenue-greenhouse-gas-emissions/#_ftn4
4- The environmental footprint of data centers in the United States. From “The environmental footprint of data centers in the United States” by M. A. B. Siddik, A. Shehabi, and L. Marston, 2021, Environmental Research Letters, 16(6), 064017. https://iopscience.iop.org/article/10.1088/1748-9326/abfba1
5- Kim, J. (2023, November 22). Four ways AI is making the power grid faster and more resilient. MIT Technology Review. https://www.technologyreview.com/2023/11/22/1083792/ai-power-grid-improvement/
6- Lam on behalf of the GraphCast team, R. (2023, November 14). GraphCast: AI model for faster and more accurate global weather forecasting. Google DeepMind. https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/
7- Prasad, K. (2023, November 17). Achieving a sustainable future for AI. MIT Technology Review. https://www.technologyreview.com/2023/06/26/1075202/achieving-a-sustainable-future-for-ai/
8- Smith, B. (2020, July 23). Microsoft will be carbon negative by 2030. The Official Microsoft Blog. https://blogs.microsoft.com/blog/2020/01/16/microsoft-will-be-carbon-negative-by-2030/
9- Xu, T. (2022, July 7). These simple changes can make AI research much more energy efficient. MIT Technology Review. https://www.technologyreview.com/2022/07/06/1055458/ai-research-emissions-energy-efficient/
