[AI Paradox] How to Use Generative AI in an Environmentally Friendly Way?

HBR Analyzes the Environmental Impact of Artificial Intelligence
Large-scale Electricity and Water Consumption in AI Development and Infrastructure Maintenance
Naver Data Center Water Usage Increased by 40% Over Two Years

Generative AI possesses remarkable capabilities that could bring transformative changes to our economy and society. However, nothing in this world comes for free. Developing AI systems requires enormous energy and costs, and maintaining the infrastructure demands a large amount of electricity. During these processes, greenhouse gases are emitted, and vast quantities of water are consumed. Last month, the Harvard Business Review introduced ways to build generative AI in a more environmentally friendly manner in an article titled "How to Make Generative AI Environmentally Friendly."

Generative AI: Massive Power Consumption and Carbon Emissions

The raw material for AI is data. It collects and processes vast amounts of data to create "information" that provides value to users. The data center industry stores and manages various information and communication technology (ICT) systems for this purpose, accounting for 2-3% of global greenhouse gas emissions. Since the volume of data doubles every two years, greenhouse gas emissions are inevitably increasing.

Data center servers require water and electricity to operate computer servers, equipment, and cooling systems. In Denmark, the power consumption for these systems accounts for 7% of total electricity use, and in the United States, it accounts for 2.8%. In particular, the graphics processing unit (GPU) chips used to run generative AI consume 10 to 15 times more power than traditional CPUs.

For example, Naver, which is actively developing AI, saw the water usage of its data center in Chuncheon, Gangwon-do, increase from 90,000 cubic meters in 2020 to 98,000 cubic meters in 2021, and to 127,000 cubic meters last year. This represents a more than 40% increase in water consumption over two years.

Not only data centers but also the process of training AI models consumes large amounts of energy and emits carbon. Some studies indicate that training large language models (LLMs) like OpenAI's GPT-4 or Google's PaLM emits 300 tons of carbon dioxide. Another study estimates that the electricity and energy consumption for training generative AI models results in 626,000 tons of carbon emissions. Although there is significant variation among studies, it is clear that the creation and use of generative AI are causing serious environmental pollution.

Avoid Developing AI Independently... Adjust Models and Evaluate Value

So, what are the ways to develop AI in an environmentally friendly manner? The report advises utilizing existing AI models. Numerous language and image providers are already available in the market. Creating and training AI models requires an enormous amount of energy. While there are attempts to develop generative AI independently in Korea, it is worth considering using models already created by large corporations and big tech companies.

It is also recommended to fine-tune existing models rather than training new AI models from scratch. Fine-tuning and prompt engineering tailored to specific content consume far less energy than training a new large model from the beginning. It is also necessary to evaluate whether AI is creating significant value. For example, is it wise to use a system that requires three times more power to improve model accuracy by only 1-3%?

Furthermore, from a global perspective, it is important to consider whether generative AI should be used for natural disasters like tsunamis and earthquakes or for medical and health-related issues, rather than simply generating blog posts or entertaining stories. Reusing technology and resources, along with continuous monitoring of carbon emissions, can also reduce the negative environmental impact of AI development.