Artificial intelligence has the potential to improve efficiency and reduce resource consumption across many industries. However, the energy AI uses is significant and appears to be rising, raising major sustainability concerns. As AI becomes more powerful and widespread, organizations are finding ways to deploy AI to help improve sustainability, while researchers continue working to mitigate its harms.
The most serious harm AI poses to sustainability is its energy consumption. The most complex AI systems, such as those currently used to process language or images, rely on neural networks, a type of program that studies patterns in data to learn a given task. Neural networks typically train on a tremendous amount of existing data—to build a chatbot that can process language, engineers might create a neural network with billions of parameters and have it analyze hundreds of gigabytes of text. Processing all this data requires running powerful computers for days or weeks on end, which can use a great deal of energy. “The deep neural network models to make the AI work have been quite large. I think that a lot of news has been drummed up by stuff like ChatGPT, and we know that that has hundreds of billions of parameters in that model, and it takes millions of dollars to train,” explained MIT professor Vivienne Sze in a 2023 podcast from MIT’s Computer Science and Artificial Intelligence Laboratory. “A lot of that cost is actually from the electricity bills, and then it takes potentially even more energy to run [...] because you're going to run it many times.”
As AI research progresses, the sizes of neural networks have skyrocketed. When OpenAI released GPT-3 in 2020, it had 175 billion parameters, a massive leap from 1.5 billion parameters in the model’s predecessor, GPT-2. Though OpenAI has not published details of its latest iteration, GPT-4, experts have estimated that the model had 1 trillion parameters when it was released in 2023. And though training an AI program requires more energy than running it, as more people use AI, the electricity consumption for operating these systems could become considerable. MIT researchers have estimated that the AI systems in a fleet of 1 billion self-driving cars would use as much energy as all the data centers in the world, due to the need to continuously analyze data from cameras and sensors throughout the vehicles.
Most creators of cutting-edge AI models don’t publish information on the energy and related emissions involved in training them, so the exact energy toll of popular models is unknown. In 2022, researchers at AI startup Hugging Face, which collaborated on an open-source large neural network to process language, estimated that training the neural network consumed over 400,000 kilowatt-hours of electricity over 118 days, close to what 40 average American homes consume in a year. They also estimated that Open AI’s GPT-3, the language model that served as the foundation of the popular ChatGPT system, emitted up to 500 metric tons of carbon dioxide during its creation, about the yearly emissions of 100 passenger cars.
As AI becomes more widely used and models continue to grow, many researchers are calling for more transparency and awareness of the technology’s environmental costs. Though researchers at Stanford have released a tool for engineers to calculate the environmental impact of the AI models they create, major AI developers like Google, Meta, and OpenAI reveal very little information about the training process used to build their models and how much energy they consume.
By learning patterns from data and using them to guide decision-making in complex environments, AI systems can help increase the efficiency of applications in many fields, leading to reductions in emissions and energy usage. Whether these benefits can offset AI’s negative environmental impacts remains to be seen, but under the right circumstances, the technology has become a powerful tool for increasing sustainability.
Using AI to analyze data on building usage and temperature—including looking at historical trends to figure out when a building is most used and when it’s hot or cool—enables organizations to reduce energy consumption in facilities. In 2022, Stockholm school building operator SISAB used AI tools from Schneider Electric and startup Myrspoven to reduce heating during high-occupancy winter hours when there was more body heat and to precool schools before peak summer occupancy, preventing temperature spiking. This led to a 15% reduction in energy usage and a reduction of 205 metric tons in CO2 emissions in the first 5 months of implementation.
Notably, AI can reduce the energy needed to cool data centers, the computing facilities where many AI systems are trained. In 2021, data center company Evoque started using AI tools to manage cooling in a 194,000 square foot data center. Sensors throughout the facility gathered temperature data, which the AI system used to understand airflow patterns and identify specific hot and cold spots. The AI found ways to keep active computers at optimal temperatures without keeping all air cooling systems on at all times, improving the facility’s energy efficiency by 20%.
AI can also make electrical grid systems more efficient by predicting where power will be needed. For instance, Nokia’s AVA software learns usage patterns of cell towers, enabling grid operators to power down specific equipment when demand is low. After implementing this software, Japanese service provider KDDI reduced power consumption by up to 50% in low-traffic environments. Applying similar technology to entire power grids, reducing energy usage by routing electricity more efficiently, is an active area of AI research.
Other industries can use AI to improve efficiency and directly reduce emissions. Shell uses AI to cut emissions during the extraction of liquid natural gas. If natural gas pressure gets too high, it is released and burned immediately, in a process called “flaring.” By applying AI to regulate natural gas extraction and minimize flaring, Shell was able to reduce yearly carbon dioxide emissions at one facility by 130 kilotons, the equivalent of taking 28,000 cars off the road for a year.
AI can also benefit sustainability initiatives directly, by improving processes like renewable energy, recycling, and environmental protection.
Renewable energy sources such as solar and wind power depend on complicated natural systems, which AI systems can analyze to use more effectively. Startup Glint Solar’s AI system analyzes satellite images and historic data to find optimal locations for solar panels. Vestas Wind Systems uses AI to analyze wind patterns and turbulence, helping them angle wind turbines so they don’t interfere with each other.
Recycling facilities have to divide mixed waste into different kinds of recyclables, which has required human workers to manually sort individual bottles, bags, and boxes. Evergreen Recycling, the nation’s leading producer of recycled PET plastic, uses AI in its facility in Ohio to speed up this task. Cameras monitor the plastic waste moving along a conveyor belt, and AI software identifies different types of plastic by color and opacity. The AI guides a robotic arm to pick up individual pieces of plastic and sort them appropriately. Using AI has doubled how many PET bottles the facility sorts per minute, and it improved the accuracy of sorting and the purity of the recycled plastic.
AI can also help analyze environmental data, monitoring ecosystem health or understanding natural disasters. Researchers at IBM have worked with NASA to develop an AI model for analyzing satellite imagery of the environment. The goal is to create a “foundation model”: an AI system that can be applied to many related tasks. Just as foundation models for language processing, like OpenAI’s GPT-3, can analyze language in many different ways, IBM’s model could analyze satellite images to help monitor deforestation or predict the effects of climate change. NASA is currently testing the model on identifying how different regions have historically flooded. “We've seen about a 20% increase [in flood mapping accuracy] in the past three months, which is really fast for how we typically do science,” said NASA Chief Science Data Officer Kevin Murphy at an IBM event.
Wildlife Insights, a project from Google, uses AI computer vision to analyze data from trail cameras to identify animals. “We processed 100 million images through the model last year alone,” Sara Beery, an MIT professor in AI who worked on the system, tells MIT Horizon. The system, she says, is “used for wildlife monitoring, understanding shifting species distributions, looking at poaching and human-wildlife conflict… there are lots and lots of different end applications for the same model, because it's universally useful.”
