What is Sophia’s doctoral thesis about?
- In my research, I investigated the true environmental impact of AI, considering not only CO2 emissions, but also the entire life cycle of its hardware and of 16 other environmental impact categories.
- Along with my co-authors, I disassembled one of the most widely used graphics processing units (GPUs) for AI training — the Nvidia A100 — and had it analyzed chemically in a specialized recycling laboratory in France. The hardware contains 32 different elements, most of which are heavy metals with toxic properties.
- Using this elemental composition data, I assessed the environmental impacts of training AI models such as GPT-4.
- A second analysis that translates computing power into hardware needs shows that training GPT-4 alone required the equivalent of approximately 2,500 GPUs and almost four tons of raw materials.
Main takeaways
- A life cycle assessment shows that electricity use causes most of the climate-related impacts, while producing the hardware is responsible for most harm to human health and depletion of mineral resources.
- The AI-related burdens are distributed unevenly across the globe. Mining, chip production, and e-waste primarily affect people in the Global South, while the benefits of AI are concentrated in the Global North.
Sophia's message
"AI is not as weightless and intangible as it may seem. Behind every chatbot are resource mines, factories, and e-waste sites, and we can measure their environmental impact beyond carbon emissions. The question is no longer how or where AI affects the environment, but rather, what policymakers, industry leaders, and users will do about it".
