Generative AI: Examining the Environmental Costs

The rapid rise of generative artificial intelligence (AI) has sparked considerable debate over its environmental costs, a prominent topic at a global summit in Paris held on February 10-11. Research conducted by the University of California Riverside and the University of Texas at Arlington highlights that GPT-3 consumes approximately half a liter (one pint) of water to produce 10 to 50 responses, underscoring the ecological concerns associated with AI technology.

Energy Consumption: A Comparative Analysis

Each interaction with OpenAI's ChatGPT, a leading generative AI chatbot, expends 2.9 watt-hours of electricity. This is tenfold the energy needed for a typical Google search, as reported by the International Energy Agency (IEA). With OpenAI claiming 300 million weekly users submitting a billion requests daily, the cumulative energy impact is significant. Beyond ChatGPT, generative AI tools are pervasive, with a survey indicating that 70% of French youth aged 18 to 24 and 65% of American teens aged 13 to 17 are engaged with these technologies.

Escalating Power Demand

Generative AI's functionality hinges on data centers, which accounted for approximately 1.4% of global electricity consumption in 2023, according to Deloitte. Anticipated AI investments suggest this could rise to 3% by 2030, equating to 1,000 terawatt-hours (TWh)—a consumption comparable to France and Germany combined. The IEA projects an over 75% increase in data center energy demands by 2026, reaching 800 TWh. Gartner, an American consultancy, cautions that this burgeoning power necessity might lead to electricity shortages at 40% of data centers by 2027.

Carbon Emissions from Training Models

Training large language models (LLMs) like those behind chatbots is another environmental consideration. A 2019 study by the University of Massachusetts Amherst estimated 300 tons of greenhouse gases emitted, equivalent to 125 round-trip flights from New York to Beijing per training session. By 2021, Oxford University researchers adjusted this figure to 224 tons for OpenAI's GPT-3. Despite these estimates, precisely assessing AI's total carbon footprint remains difficult, due in part to limited transparency in model production and the absence of universal measurement standards.

Water Usage for Cooling

The environmental impact extends beyond energy consumption, as substantial water usage is required for cooling AI hardware. GPT-3 uses roughly half a liter per 10 to 50 responses. Increased AI demands are projected to consume between 4.2 billion and 6.6 billion cubic meters of water—four to six times Denmark’s annual water consumption, according to a 2023 study.

Growing Electronic Waste

Generative AI applications contribute significantly to electronic waste, with an estimated 2,600 tons generated in 2023, as reported by the journal Nature Computational Science. Without intervention to curb waste, this might reach 2.5 million tons by 2030—akin to 13.3 billion smartphones discarded. The required hardware often necessitates rare metals, primarily sourced through environmentally damaging mining processes across Africa.