AI Contributes to Climate Change—Here's How to Make It Cleaner Technology

Generative artificial intelligence (AI) is all the rage, but producing so much information has an environmental cost. 

A recent report by Stanford University researchers finds that just training the model behind the AI chatbot ChatGPT released emissions equivalent to those of 9 cars. While the exact amount of AI's environmental impact has yet to be measured, experts say the consequences are severe. 

"We need to continue improving both hardware and software," tech advisor Vaclav Vincalek told Lifewire in an email interview. "More efficient data centers, better cooling systems for those data centers, energy-efficient algorithms, and incentives to pursue innovations that can help AI developers reduce the environmental impact of their systems."

The Environmental Impact of AI

The Stanford research examines the carbon costs associated with training four models: DeepMind's Gopher, BigScience initiatives' BLOOM, Meta's OPT, and OpenAI's GPT-3. OpenAI's model reportedly released 502 metric tons of carbon during its training, four times more carbon than Gopher and 20.1 times more than BLOOM. GPT-3 used the most power at 1,287 megawatt-hours.

"There have been plenty of reports about the environmental 'cost' of training generative AI models like ChatGPT," Vincalek said. "You'll read one paper say it consumed as much energy as 120 homes in the US would consume over the course of a year. Another report likened it to spewing emissions equal to a single person flying between New York and San Francisco 550 times a year. Then there's water consumption associated with generative AI training. One report says it guzzled 185,000 gallons of water over the course of its training."

The problem is one of scale. For generative chatbots to be able to provide answers, they require 'training' on very large datasets, noted tech consultant Sam Cooper in an email. This training process requires the intensive use of powerful supercomputers and processors, which use a significant quantity of electricity. For example, it took nine days to train one of OpenAI's early AI chatbots, consuming over 25,000-kilowatt hours of energy. This amount is equivalent to the energy used by three US homes for an entire year.

"It isn't just the energy consumption that is an environmental concern for AI," he added.
"The processors and chips used by the supercomputers to train these AI models require large quantities of silicon, plastic, and copper."

The increasingly competitive nature of the AI business also contributes to environmental issues, Hammad Khan, the CEO of AlphaVenture, an AI consulting firm, said via email. He said the AI arms race is leading to a surge in the demand and production of chips with a significant carbon footprint.

"Also, with sudden demand spikes and rapid advancements, you often end up with piles of outdated hardware in junk, as we have already witnessed in the crypto boom," he added. 

But AI Can Also Help Reduce Climate Change

It’s not all bad news when it comes to AI and the environment. The technology has the potential to significantly reduce waste and promote sustainability by facilitating the repurposing of old or broken items, Jake Maymar, the vice president of innovation at the tech consulting company The Glimpse Group, said in an email. 

AI can automatically sort and categorize various types of waste, such as electronics, textiles, and plastics, based on their composition and condition, Maymar noted. "This can enable efficient identification of reusable or repairable items that can be repurposed instead of being discarded as waste,” he said. 

Also, AI can assist in optimizing the design of products to make them more durable, repairable, and reusable. “Using AI algorithms, designers can optimize the use of materials, improve product durability, and reduce waste in the design phase itself,” Maymar said. 

Ultimately, the AI industry must find cleaner ways to produce data. Engineers need to prioritize energy efficiency in designing the hardware and software used for training and inference, Krzysztof Sopyla, the head of machine learning at STX Next, said in an email. This approach could include using more energy-efficient processors and algorithms that require fewer calculations.

"Another approach is to use renewable energy sources to power the computing infrastructure used for generative AI,” Sopyla said. “Many data centers are already making progress in this area, and I believe this trend will continue to grow in the future.”