Artificial Intelligence and the Environment

---

• Energy intensive data centers used to power AI generate a large amount of heat, requiring constant cooling using freshwater. At the end of this cooling process, water is significantly contaminated with chemical coolants and heavy metals and discarded.
• Data centers supporting GPT-4 in Iowa were reported to have consumed up to 6% of the district’s water in the last month of training the AI model.

• There is now increasing demand for rare earth materials, such as lithium, cobalt, zinc, needed to build the specialized equipment required for AI development.
• Extraction of these rare earth minerals have been linked to severe negative consequences, such as deforestation, soil erosion, groundwater contamination, and biodiversity loss and social impacts such as human rights violation and political instability, particularly in vulnerable regions.

• In 2022, 62 billion kg of e-waste (electrical and electronic equipment waste) was produced around the globe (a new record). Only 22% of that amount was recycled.
• E-waste can impact communities through multiple ways including lead and mercury emissions and leakage of plastic in ecosystems. Currently only 42% of nations have policies in place to regulate e-waste, and even fewer with collection targets.

So… what can we do to mitigate these dangers?

• Explore alternatives to AI – use AI in specific-use cases where no alternatives exist or opt for human decision making and avoid using generative AI tools for simple tasks such as a conventional search on the web.
• Conduct a cost-benefit analysis to determine if the environmental cost of the AI application outweighs the environmental benefit gained (see here for some criteria).

• Manufacturers of computational hardware should provide details around the environmental impacts of their assembly process, including carbon emissions, water usage, and rare earth materials used, to help increase transparency in the AI supply chain.
• Computer scientists and researchers should disclose, to the best of their ability, resources leveraged in training and deploying AI models (for example, time, carbon emissions) to better enable comparisons between models.
• Separate AI emissions from other ICT emissions, for example dedicated data centers to train AI, in order to get access to more accurate AI carbon footprint data.

• Reuse water consumed in data centers (e.g filter and reuse water used in cooling servers) and use alternate sources other than fresh drinking water (for example treated wastewater).
• Dismantle, reclaim, and recycle materials from processors and old lithium batteries in or around data centers to help build a closed loop on-site whilst also reducing transportation emissions.
• Aim to use at least 50% of recycled/re-used materials when replacing parts or building future data centers to reduce E-waste.
• Use clean and low-tech energy to power data centers.

---

• Increased efficiencies brought about by AI are lowering costs of producing goods and services. While it can lead to positive impacts at the individual product level, these benefits can be undone by overall increases in demand for the product.

• AI advancements can help identify new extraction sites for carbon intensive industries such as fossil fuels, adding to further environmental degradation and fueling greenhouse gas emissions.

• Given its cross-cutting nature, it is extremely difficult to map the growing number of indirect pathways by which AI impacts our environment. This uncertainty is troubling, as there is a growing concern that indirect risks of AI could outweigh the direct impacts.

1. Encourage policy makers to adopt a precautionary principle and design policies that acknowledge, anticipate, and limit the negative indirect impacts of AI.
2. Awareness around the environmental impacts of AI should be increased across policy makers, youth, tech innovators, civil society, and researchers to help develop AI tools that center around environmental and social values.
3. Increase access to education programs on environmental digital sustainability, for example, a mandatory certification on digital sustainability for ICT sector professionals and CEOs and other individuals in leadership positions at high stake AI companies.
4. Scenario mapping and quantification: Researchers (and climate modelers) are calling for a better understanding of how AI can impact climate goals under different socioeconomic and political settings, with some proposed starting points.

---

---