ChatGPT makes a splash with AI’s water footprint

Sarah Gabriel, Production Assistant

ChatGPT has raised concerns about not only its advances in artificial intelligence but also its impact on the environment.

A new study, titled “Making AI Less ‘Thirsty’,” by the University of Colorado Riverside and the University of Texas Arlington projected that ChatGPT-3 training consumed 185,000 gallons of water. The total is “equivalent to the amount needed to fill a nuclear reactor’s cooling tower,” according to Gizmodo.

Developed by OpenAI, the chatbot answers questions that its users input by using a large language model, or LLM, algorithm. LLMs utilize large sets of data collected from the internet, such as web pages, scientific research and social media posts, to help understand and predict content.

Microsoft partnered with OpenAI to train its AI models using supercomputers that contain 10,000 graphics cards and over 285,000 processor cores, according to Interesting Engineering.

These numbers helped the researchers estimate the scale of operations the AI models are working on.

To put this scale into perspective, a discussion with ChatGPT of about 25 to 50 questions would consume 16.9 ounces of water — equivalent to a single-use bottle.

Although online functions such as sharing, uploading and using ChatGPT happen digitally, the actual data is physically stored in large data centers. These data centers generate large amounts of heat which require cooling systems to protect the equipment from malfunctioning.

Evaporative cooling towers are often used to help chill these data centers and consume large amounts of water.

A video by SPX Cooling Technologies showcased the cooling process and explained that cooling towers have an evaporative side and a condenser side.

In the condenser side, the chiller absorbs heat from the evaporator and sends warm water to the cooling tower. This warm water is chilled by evaporative cooling and then returned to the condenser to pick up more heat.

The evaporator side collects warm air, cools it and then routes the air to data equipment to keep it cool. Most water consumption is a result of this evaporative process.

“Around a gallon of water is consumed for every kilowatt-hour expended in an average data center,” Mack DeGeurin of Gizmodo said.

Water consumed in this process must also be clean freshwater in order to avoid corrosion and bacteria growth.

Aside from cooling systems, data centers also require a high production of electricity that scientists call “off-site indirect water consumption,” according to Interesting Engineering.

Absurd water consumption is not limited to ChatGPT or other AI models. But with the increase of generative AI models such as Google’s Bard AI, researchers have drawn their attention to AI’s water footprint.

Climate change and preexisting droughts also amplify concerns over water usage.

“Figures suggest 2.2 million people in the United States are without running water and basic indoor plumbing; more than 44 million people have inadequate water systems,” the World Economic Forum said.

The demand for water is rising as technology becomes increasingly integrated into society.

“Water footprint must be addressed as a priority as part of the collective efforts to combat global water challenges,” researchers of the study said.