Generative AI’s Dirty Secret

The swift expansion of artificial intelligence (AI), namely, generative AI and large-language models (LLMs), has caught many people off guard, but the effects on the environment continue to be a major concern. This surge has raised concerns about the strain on electrical systems due to the increasing electricity consumption of data centers. Large language models require tens of thousands of sophisticated high-performance computers to handle and analyse enormous volumes of data. Once trained, these models can make predictions about fresh data and answer questions. Specialized electrical circuits called graphics processing units (GPUs) are widely used because they can perform many calculations or processes simultaneously, but they use more energy than many other types of processors do. One problem is that the carbon footprint of artificial intelligence has not received enough attention. While there are growing initiatives to incorporate renewable energy supplies, the majority of data centers currently function around the clock and are powered by conventional energy sources. Data centers are responsible for 2.5--3.7% of worldwide greenhouse gas emissions due to their energy consumption, which is greater than that of the aviation sector. Nearly one hundred million people were using ChatGPT a few weeks after it launched. Many individuals are keen to use AI for every purpose instead of using the current web searches that depend on simpler AI models; however, according to one tech professional, a single ChatGPT query could consume 100 times as much energy as a single Google search. Researchers discovered in 2019 that the energy required to create a generative AI model with 110 million parameters, named BERT, was equivalent to one person’s round‒trip transcontinental flight journey. The size of any AI model is indicated by its number of parameters; larger models are typically more sophisticated. According to research, the construction of the much larger GPT-3, which has 175 billion parameters, required 1,287 megawatt-hours of electricity and produced 552 tons of carbon dioxide equivalent, or the same amount of carbon dioxide created by 123 passenger cars running on gasoline for a year.

In the last few days, adorable, whimsical AI-generated images that imitate the classic Studio Ghibli style have taken over social media. Millions of people are now able to create images in ChatGPT owing to OpenAI’s features. AI fans soon found that GPT could imitate popular art styles, such as the beloved artwork created by the animation powerhouse Studio Ghibli, because it could render graphics in a variety of art styles. When AI is used for creative jobs, such as creating Ghibli-style images, there are important and frequently disregarded environmental effects. Discussions concerning the sustainability of these technologies and their long-term environmental implications have arisen as a result of the rapid surge in popularity of these AI-generated photos. As reported by OpenAI CEO Sam Altman, the anime-mimicking image generator burned the company’s GPUs, and it was so popular that OpenAI postponed a rollout for free ChatGPT customers. This speaks volumes about the appeal of AI, but it may reveal even more about how sustainable that popularity is.

AI has implications that go beyond carbon emissions. Water usage and the production of electronic trash from data centers, which are frequently powered by fossil fuels, are two examples of how AI affects the environment. The cooling of all those data centers is also necessary. Many data centers that use AI cloud models choose liquid cooling because of the high power requirements of AI. Additionally, GPUs may still melt even in that case. To keep its data center cool, OpenAI requires more than two litres of water for every fifty queries as of October 2024. In addition to the use of a large amount of water, the water cycled through by liquid cooling evaporates because such data centers are very hot. To reduce the environmental impact on freshwater supplies, data centers are increasingly adopting saltwater. Conversely, the adoption rate is still not 100 per cent. By 2030, Google wants to achieve net-zero emissions. The goal of reaching net zero is to promote innovation, develop new solutions, and try out various strategies rather than to slow down or compel companies to comply.

The carbon footprint of AI highlights the urgent need for knowledge of clean electricity, grid management, decarbonization, and carbon removal—knowledge that will become increasingly important as more businesses recognize the difficulty and expense of the path ahead. It is important to consider the environmental impacts of artificial intelligence, as society accepts its advantages. The United Nations Sustainable Development Goals (SDGs) are significantly hampered by the carbon footprint of artificial intelligence (AI) technology, especially in regard to SDGs 12 (Responsible Consumption and Production) and 13 (Climate Action). To ensure that the development of AI does not come at the expense of the world, innovation and sustainability must be balanced. The discourse surrounding the carbon footprint of artificial intelligence is still in its nascent stages; however, it compels us to reassess our utilization of AI in both professional and recreational contexts.

Bitan Misra is an Assistant Professor, Nilanjan Dey is an Associate Professor, both at Dept of Computer Science & Engineering, Techno International New Town, Kolkata. Views expressed are personal