Arguably no other tech is as hyped or provoking the same level of fascination as ChatGPT and other versions of generative AI.
Indeed, generative AI (GenAI) has exploded onto the tech scene and into public consciousness. While the tech has a number of benefits, it’s worth spending some time learning about its sustainability impact in order to make better decisions.
Like many emerging digital technologies, it's negatively affecting the environment in more ways than most people realize. And with pressure growing to monitor carbon emissions and report on environmental, sustainability and governance (ESG) issues, companies cannot afford to ignore this important issue, especially since GenAI has high resource requirements and a hefty carbon footprint.
"The amount of computing required to run [GenAI] models is huge," said Ravi Jain, chair of the US Technology Policy Committee Working Group on Generative AI at the Association for Computing Machinery (ACM). "And when you run all these computers trying to run these models -- whether to train them or use them -- they require tremendous amounts of energy."
That translates into a spike in carbon emissions, which adds to the total from digital technologies, he said.
"In the U.S. we expect the amount of energy being used will double or triple in next couple of decades, not just because of GenAI, but because of all the digital technologies we're using," Jain said. "And that increase is occurring faster than we can supply it with renewable energy."
GenAI, large language models and energy requirements
As with many other types of AI, a major environmental concern about GenAI is due to its high energy requirements.
Indeed, researchers have voiced concerns about AI's environmental impacts. But they've called out GenAI specifically because its energy requirements are even more significant than other types of intelligent technologies.
Ravi JainChair of the US Technology Policy Committee Working Group on Generative AI at the Association for Computing Machinery
To understand why generative AI is particularly compute intensive, it's important to understand how it works, Jain said.
Generative AI is a type of AI designed to create -- or generate -- content such as text, images, video, audio and computer code. GenAI systems feature large statistical models and are trained on huge data sets. The models tell the systems how to analyze data to find patterns, while the training teaches the AI the information it needs to produce content when asked to do so.
Inference happens when those trained AI systems go to work, taking what they've learned to analyze new data and generate content or predict outcomes in response to a user's query or prompt for information.
From an environmental standpoint, both training and inference require significant computational power, which in turn requires significant amounts of electricity to run. As the models become larger, and the training data and parameters become larger, more computational power is required and, thus, more electricity is needed.
"Over the last couple of years, the main focus of vendors for GenAI has been to increase performance and accuracy," explained Reece Hayden, a senior analyst at ABI Research and lead of the firm's AI and machine learning research service. "When vendors drive accuracy in AI models, they increase the parameters, or size, of the models; that means more time [running computers] and more energy."
GenAI's inefficient resource utilization only exacerbates the problem, he added.
The carbon emissions of GenAI and large models
The computational work to run Generative AI models such as ChatGPT, Bard and Claude occurs in large cloud computing data centers.
Cloud computing companies have been focusing in recent years on making their data centers more sustainable by using more renewable energy sources and locating facilities in climates optimal for operations. Nonetheless, cloud computing still leaves a notable carbon footprint: The International Energy Association estimated that data centers accounted for 1% of energy-related global greenhouse gas emissions and approximately 300 metric tons of carbon dioxide equivalent in 2020.
And that number is growing quickly as society increasingly relies on digital technologies.
Meanwhile, some researchers have calculated the carbon footprint of typical AI training models.
Training a single AI model can emit 626,000 pounds of carbon dioxide equivalent, according to a 2019 study "Energy and policy considerations for deep learning in NLP," published by the University of Massachusetts, Amherst. That figure is nearly five times the lifetime emissions of the average American car.
GenAI: The hungry, thirsty tech
Generative AI has been linked to other environmental issues beyond its carbon footprint.
The technology also requires more hardware than other types of computing, Hayden said. And it cycles through that hardware faster, meaning the replacement cycle is also more intensive than in other types of computing.
Karen PanettaFellow with the Institute of Electrical and Electronics Engineers
Those hardware requirements add to a manufacturing process that already stresses the environment, Jain said. The required hardware uses rare earth elements that must be mined and transported.
More e-waste is also a byproduct of generative AI.
The shortened lifespan of equipment used in GenAI adds to the world's growing amount of electronic waste, said Karen Panetta, a fellow with the Institute of Electrical and Electronics Engineers.
"We do not do a good job today with the recycling of computer chips and elements, and we don't do a good job today at taking back copper, said Panetta, who is also a professor of electrical and computer engineering at Tufts University and dean of graduate education for its School of Engineering. "That is a problem that is pervasive now, and it will continue to be a pervasive problem because we're not addressing it."
Why companies overlook generative AI concerns
Fears of being left behind have business and tech leaders racing to discover ways to use ChatGPT and other GenAI tools. Few stop to think about their sustainability and environmental impacts.
For executives interested in using GenAI, sustainability is not their top concern, Hayden said. On a general level, they're primarily interested in understanding the technology and how it will impact their industry, company and market. They're also studying how to use the technology to differentiate their products and to create efficiencies within their operations -- all while adhering to their security and privacy standards.
On a more specific level, trustworthiness is probably the main concern business leaders focus on when making decisions about generative AI projects, Hayden said. They realize they need to solve issues such as hallucinations and bias. On top of this, performance, time to deployment and cost are important as enterprises look to build out the business case and assess the feasibility of deployment.
"I expect sustainability will have a larger impact [on enterprise decision-making] moving forward."
As interest in sustainability grows, attention to generative AI's environmental impact may grow as well.
Some developers do include carbon estimates in model cards, the files that accompany AI models which explain their intended use, limitations, evaluation and parameters, Jain said. For example, the model card template from the open source HuggingFace is available at the software development platform GitHub. ACM's own guidelines for the development, deployment and use of GenAI highlight the increasing attention paid to the subject of sustainability.
Still, the real environmental impact of generative AI is difficult to pin down, as most generative AI companies -- such as the popular OpenAI -- keep a tight lid on information about electricity consumption, specialized chips needed and so on, although researchers are working to make estimates using information they can glean.
For example, the increase in demand for AI chips is one basis for a somewhat conservative prediction that by 2027, AI servers could use between 85 and 134 terawatt hours each year, a number almost as high as what Argentina uses each year, according to a new study, "The growing energy footprint of artificial intelligence."
Tips for reducing GenAI environmental harm
Industry experts and organizations are giving more attention to the issues surrounding technology and climate change so it pays to follow new research and thinking about ways to lower the negative environmental impact of AI model development.
Business and tech stakeholders can follow these tips for creating more environmentally friendly generative AI now, according to Panetta, Jain and Hayden:
- Choose renewable energy. Use vendors that source electricity from renewable sources to the greatest extent possible.
- Choose existing models. When possible, use existing generative models, fine-tuning them to meet needs instead of creating entirely new ones, and reuse models and resources when possible.
- Don't overtrain. Avoid overtraining AI models, and use only the data required to meet use case needs. Not all use cases require the same level of accuracy from the models, so training to the highest possible degree of accuracy when it's not required may not be worth the environmental cost.
- Think small. Instead, right-size the models required and use the smallest possible models because they're the most energy-efficient.
- Conserve energy. Use energy-conserving computational methods.
- Study model cards. Use model cards to help make decisions and incorporate the carbon footprint of your AI activities into the company's carbon monitoring program.
- Question whether the project really needs GenAI. Be selective in the use of GenAI, and opt for less energy-intensive solutions when appropriate.
GenAI to support sustainability
Generative AI's largely negative environmental impact may lessen as the technology becomes more sophisticated and efficient.
Technology and design advances could help GenAI become greener, Panetta said. In general, predictions about how GenAI will impact the environment don't account for those advancements.
"People are always assuming we're going to use the things in the exact same way, but we have to look at the advances and new materials coming out," she said.
Low-power electronics, systems designed for efficiency, more efficient utilization of compute resources, computer code written for efficiency and more energy-efficient computer chips could all help lower the environmental impact of GenAI, Panetta said.
Generative AI could even be used to help sustainability and ESG efforts.
Certain GenAI uses have the potential to create more environmentally friendly processes, said Randal Kenworthy, senior partner at digital services firm West Monroe. Such uses may be able to offset the technology's environmental impact. For example, the technology may help organizations calculate their carbon footprint and other environmental impacts, support the development of impact reduction strategies and help devise design and production processes that have lower environmental impacts. Mid-market organizations may especially see big benefits as they typically haven't had the resources to address these environmental concerns to the degree that larger enterprises have.
"GenAI streamlines a lot of that [ESG] work," Kenworthy said. "It reduces the manual efforts, and I think that's where it's going to have a transformative impact."