The Hidden Environmental Costs of Artificial Intelligence: A Call for Sustainable Practices

A recent report from the United Nations University Institute for Water, Environment, and Health (UNU-INWEH) has revealed alarming predictions about the environmental impact of artificial intelligence (AI) by the year 2030. The findings highlight that the water consumption associated with AI could match that of 13 billion people in sub-Saharan Africa, while the energy required may reach three times the annual consumption of Pakistan, Bangladesh, and Nigeria, which together have a population of around 650 million. Furthermore, the carbon emissions generated by AI could soar to 400 million tonnes of CO₂ equivalent—an amount comparable to the total emissions of the United Kingdom. These projections are compounded by the fact that the operational footprint for AI will demand 14,500 square kilometers of land, which is twice the size of the Jakarta metropolitan area, home to over 32 million residents. The report emphasizes that if data centers powering AI were treated as a country, their current electricity consumption of 448 terawatt-hours (TWh) would be equivalent to France's total electricity usage. It warns that the environmental costs of AI are consistently underestimated, focusing primarily on the carbon emissions that stem from the training of AI models. This perspective obscures the broader implications including significant water usage for cooling and energy generation. Professor Kaveh Madani, director of UNU-INWEH, stated that the report is not intended to denounce AI, but rather to urge responsible application and proactive solutions to address its unintended impacts. He adds, "We have a narrow window to ensure that the backbone of the technological revolution of our era develops within planetary limits." The study also reveals that the energy consumption of AI is dominated by inference—the operations conducted when users interact with AI models. These inference processes account for 80-90% of total energy usage, a stark contrast to earlier beliefs that training was the primary energy sink. When looking at specific applications, the report demonstrated that a standard chatbot conversation can consume 200 times more energy than simple tasks such as classifying emails. Complex operations, like generating synthetic images or short videos can amplify energy consumption even further—up to 200,000 times more. The conclusions drawn from this report indicate a troubling trend: the benefits of AI and the negative externalities it produces are distributed inequitably. For instance, in Ireland, data centers accounted for 21% of total energy consumption in 2023. This excess has resulted in halted constructions of new data facilities in Dublin. On the other hand, countries like Uruguay are experiencing infrastructure developments for water-intensive data centers amid drought conditions affecting local water supplies. Moreover, the report predicts that by 2030, AI infrastructure will contribute to approximately 25 million tonnes of electronic waste each year. This waste, primarily from obsolete processors, is likely to accumulate in low-resource countries, exacerbating already existing challenges in those regions. Despite the concerning findings, the report emphasizes the need for greater transparency regarding AI's environmental footprint and encourages policy recommendations for governments and developers alike. The report advocates for standardized reporting on the environmental impact of AI operations, along with efficient design practices that prioritize model selection based on task complexity. The UN's call for increased scrutiny on the environmental footprint of AI marks an important step for sustainable engagement with this emerging technological landscape. As the conversation around AI continues to grow, understanding its environmental implications must remain a paramount concern—ensuring that technological advancements do not come at the cost of our planet. Related Sources: • Source 1 • Source 2