ChatGPT – a water-guzzling giant?

The fashion for ChatGPT and AI in its various applications and dimensions has taken over the world. While the use of ChatGPT for writing school essays or blog content is controversial, the use of artificial intelligence in science, medicine, industry or agriculture raises great hopes for streamlining processes, saving resources and improving quality. However, the proverbial medal has two sides, and so does AI. The dark one is environmental costs. Water consumption during GPT-3 training is compared to enough to fill a nuclear reactor with water. It seems impossible. Are you sure?

ChatGPT

ChatGPT (Generative Pre-trained Transformer), developed by OpenAI, is a language model whose sophistication and accumulated datasets make it possible to generate extended text/responses in different domains and on different topics. Launched a year ago in prototype form, it gained more than a million users in just five days, and the number keeps growing. Artificial intelligence is constantly learning to meet their expectations. This seemingly purely practical and useful tool also has its drawbacks, including a significant environmental impact.

Water and AI development

The huge demand for AI-supported solutions comes with an equally huge cost to “produce” artificial intelligence. This does not only apply to technological costs, but also to increased water consumption. This year, with the 2022 published. by Microsoft’s environmental report, the world was circulated with the news that the corporation’s water consumption had increased by as much as 34 percent. (to nearly 1.7 billion gallons) in 2021-2022. Such an amount would be enough to fill 2,500. Olympic swimming pools. The surge came despite Microsoft’s pledge to take a holistic approach to data center water consumption, among other things. through the use of closed circuits and full process control. The situation is linked to the establishment of a partnership with OpenAI.

A similar circumstance is expected at the other industry giants. The Google report, in support of these assumptions, finds a 20 percent increase in water consumption. Building a suitable language model, based on human-generated data sets, requires the consumption of a lot of electricity, but also water. Especially in hot weather, its intake increases because it is needed to cool the supercomputers in the data centers. At the same time, it should be noted that this is only one element affecting the size of AI’s water footprint, the other being indirect consumption – for electricity generation.

Development on the river

The 50-mile-long Raccoon River drainage basin, in central western Iowa, is part of the Mississippi River basin. The intake located on it supplies drinking water to residents of the nearby city of West Des Moines. There would be nothing unusual about this if it weren’t for the concerns of local authorities about ensuring a continuous water supply, and not because of pollution or drought, but because of Microsoft’s investment in AI.

In May 2022. The company revealed that it has built an “advanced AI supercomputing data center” to train the GPT-4 OpenAI model. The choice of location was not accidental – temperatures in this region typically range from -9°C (cold and snowy winters) to 30°C (warm and humid summers). Water-based cooling is necessary only when the temperature exceeds 29°C. Whether the West Des Moines authorities’ concerns are justified time will tell. This July, temperatures exceeded 36°C in the city.

AI water consumption estimates

Despite sparse data on the development of artificial intelligence and general information from reports on the sustainability of AI companies, a team of researchers from the University of California, Riverside (Shaolei Ren et al.) has attempted to estimate the water and carbon footprint of AI models – GPT. The research was based on information on 10 geographically dispersed data centers. Based on these results, it was estimated that ChatGPT consumes about 500 ml of water per interaction consisting of 20 to 50 questions.

These estimates include both direct and indirect water consumption. In contrast, water consumption just for GPT-3 chat training in data centers was estimated at 700,000. l. This amount would be enough to fill the cooling tower of a nuclear reactor or produce 370 BMW cars.

The researchers stress that raising public awareness of the environmental costs and the impact of AI development, including locally, on the living conditions of local residents is essential to motivate investors to care about the sustainability of the industry. Including making the right location choices, taking into account climate change and the need for environmental sustainability.

ChatGPT – answer

Satisfying my own curiosity, I checked the answer to the question: how much water does ChatGPT use? generates AI. In doing so, I asked a supporting question. Below is its content, without making language corrections:

Servers running ChatGPT consume a significant amount of electricity and water, although it is difficult to pinpoint specific amounts because they depend on a number of factors, such as the location of the servers, the efficiency of the hardware and cooling technology, and how they are operated. For servers that process huge amounts of data, such as those supporting language models, there is a need to maintain proper temperature conditions, which may involve the use of water cooling. However, water consumption for servers is usually much lower than electricity consumption.

A company or organization managing servers can use various strategies and technologies to optimize energy efficiency and minimize environmental impact. All this can affect water consumption. It is worth noting that innovations in sustainable server infrastructure are being developed that can help reduce both energy and water consumption.

Do you have any reports on this subject?

The data in the April 2023 report exemplifies the water consumption estimates of language models such as ChatGPT-3. Of course, it’s worth noting that such estimates are approximate and can vary depending on a number of factors, such as cooling technology, server location and specific infrastructure configuration. In the case described, the 700,000 liters of water used to “train” ChatGPT-3 refers to the process of training the model, which is a computationally intensive task and can generate a lot of heat, which requires cooling. However, this is only one aspect of water consumption.

Stating that ChatGPT uses the equivalent of a 500-milliliter bottle of water to conduct a discussion with a user containing 25-50 questions is more specific and more tailored to illustrate the impact on water consumption when interacting with a model. This figure of 500 ml is probably an approximation and may vary depending on a number of factors, such as server load, cooling performance and many others.

As the Norwegian government agency points out in its June report Ghost in the Machine , the world’s demand for water will continue to grow. The technology industry, including, among others. The development and deployment of generative artificial intelligence is a sector contributing to increased demand for water. Despite the IT giants’ assurances that their goal is to achieve neutrality in both water consumption and_CO2 emissions, there is no evidence of this. Neither company has published data supporting these claims and pointing to steps by which artificial intelligence could become environmentally neutral.

It is also worth taking into account that AI development is not only carried out by corporate giants, but also by many smaller companies that, while improving AI-assisted technologies, do not take into account environmental responsibility. So the way forward, in terms of reducing the industry’s water footprint, doesn’t seem so obvious, and in fact it could be argued that it hasn’t been mapped out by anyone.

The punchline of this article was generated by ChatGPT itself, summarizing the second of the questions:

It is worth remembering that the exact water consumption figures of language models are usually closely guarded by the companies running such models and may not be readily available to the public. Approximate figures are only available based on estimates and surveys by independent experts.