Researchers seek algorithms that won’t destroy the planet: ‘ChatGPT’s energy consumption must be brutal’

One day in 2019, while reading a scientific article, a light bulb went off for researcher Verónica Bolón, an expert in artificial intelligence. “It said that training an AI language model emitted CO₂ equivalent to that of five cars over its lifetime. I was shocked. Artificial intelligence consumes a lot of energy as it learns, it’s no longer just when it is used, but until then I had never thought about it,” she recalls sitting in one of the rooms at the University of A Coruña’s Center for Research in Information and Communication Technologies (CITIC). She and seven other researchers (a total of four men and four women) form one of the teams in Spain that are currently searching for green algorithms; in other words, they are immersed in resolving the great paradox brought about by the boom in supercomputers, data traffic, the cloud and ChatGPT. “Artificial intelligence helps us to solve many problems, some of them derived from climate change, but at the same time it is part of the problem,” Bolón observes.

Green artificial intelligence strives to design solutions that achieve the proposed objectives, but in an environmentally sustainable way, that is, with the efficient use of computational resources. This concept, which is barely three years old, contrasts with red AI, which only values algorithms for their performance, even if they involve high energy consumption. “Now green algorithms are becoming more important, although not as [important as they] should be,” says Verónica Bolón. She believes that this lack of attention is due to the fact that taking sustainability into account goes against the interests of the big technology companies in the dizzying artificial intelligence race. Microsoft, Google and OpenAI, she explains, are leading the field by consuming “huge” amounts of energy. Environmentalists, the researcher points out, have been warning about the environmental footprint of this process for years. Indeed, in 2010 Greenpeace reported that the farms of thousands of servers that make up the cloud “require enormous amounts of energy to operate and cool” down.

The emergence of ChatGPT, a fascinating tool that millions of people use every day, has exacerbated the problem. “We don’t know exactly what it consumes, but it has to be [a] brutal [amount of energy]—both in training it and using it—because it needs a lot of data and very large neural networks. And it’s not something people stop to think about, because they don’t have information about it either. [The company] is called OpenAI, but in [terms of information about energy consumption] it’s not open at all,” Bolon says of the company that launched the popular app, which was founded in 2015 by Sam Altman and Elon Musk, among others, as a nonprofit organization.

A 2020 University of Copenhagen study estimated that a single training session of GPT-3 (the text generator on which ChatGPT was based) consumed as much energy as 126 Danish houses over the course of a year, Bolón says. At the moment, no energy consumption information has been published for ChatGPT-4, the next version, but there is one piece of data that gives an idea of its environmental impact: it was trained with 1 trillion parameters versus the 175 billion used with GPT-3. “I’m not saying you have to stop that technological development. But if you ask me whether, ethically, it’s worth all that computational effort for a model that is dedicated to entertaining or helping to perform tasks, I have mixed feelings. I’m impressed by the progress, but I think we’re focusing on the wrong things. It’s not being used for the fundamental issues that affect society,” she says.

In 2018, Verónica Bolón, 39, began to investigate how to run algorithms on very small devices that did not have the computational capacity of a computer. She was not yet familiar with the concept of green artificial intelligence, but soon she realized that using fewer computational resources is more sustainable because it consumes far less energy. That’s also true if data is not continuously transferred to the cloud, she explains. In 2022, the same year she joined the Royal Academy of Exact, Physical and Natural Sciences of Spain, she was awarded funding for a national research project that will last until 2025. Her team has managed to execute algorithms, without losing precision in the results, with only 16 bits (one bit is the minimum unit of information) instead of the 64 bits that common computers use today. The precise energy reduction implied by the change is not yet known because they are working with simulations, Bolón explains.

Her team—which includes Brais Cancela, Jorge González, Noelia Sánchez, Laura Morán, David Novoa, Eva Blanco and Samuel Suárez—is also exploring another path toward green artificial intelligence, which consists of speeding processes up. The researchers are working on accomplishing the same things in less time. “It’s about racking our brains to optimize the models, so that they work just as well with less data and a smaller network. For example, now everyone wants to apply deep learning, but there are problems that can be solved with much simpler models. It is [like] killing flies with cannons. We have to try to discern when it is necessary and when it is not,” says Bolón.

Curbing AI emissions by law

This scientific center has about 200 researchers, and 75% are dedicated to the field of artificial intelligence and data science. It is part of the university and business ecosystem for which A Coruña has been selected by Spain’s government to be the headquarters of the Spanish Agency for the Supervision of Artificial Intelligence (Aesia). Green algorithms will be part of the agency’s work. Since last December, Spain has had a National Plan for Green Algorithms, which aims to promote both the energy efficiency of artificial intelligence and its use for solving environmental problems. The plan includes the creation of up to two chairs in this discipline; the University of A Coruña has responded to the call for applications and decisions will be made “shortly,” according to sources from the State Secretary for Digitalization and Artificial Intelligence.

Of the €257.7 million of European Next Generation funds with which the plan is endowed until 2025, €6.9 million will be allocated to research projects in this field, the State Secretary for Digitalization and Artificial Intelligence says. “It is easy to talk about the importance of sustainability and reducing consumption in computing, but to do that we have to promote research lines and that requires funds,” says Manuel González Penedo, the director of CITIC. He calls for improvements in the structural funding of centers like the one he oversees, program continuity and less bureaucracy to attract talent from outside Spain.

Verónica Bolón advocates a law for regulating artificial intelligence from the perspectives of both ethics and sustainability: “Regulation is necessary and it must include consumption. Are we going to allow information technologies to be responsible for 20% or more of CO₂ emissions into the atmosphere? Then we will have to rein it in, as in other industries and areas.”

Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition