Abstract
Every time people ask an AI chatbot to write an email or create a summary of a piece, a whole series of processes takes place: processors become active, cooling devices come into action, and power plants consume either fuel or water using their turbines. Generative AI has already burrowed its way into our culture; however, the environmental impact that it produces remains hidden. My research is aimed at exposing this issue by examining two different sides of it: first, there will be the technical side, which will allow me to calculate the actual cost of a language model in terms of energy and water footprint. Second, there will be an STS analysis, which will investigate why the public picture of these environmental costs is so distorted.
On the technical side, my capstone explores a rather intuitive but overlooked issue: how much electricity and water is required to deploy an open-source language model with very low computing power, and which basic modifications could decrease that cost? Using open-source language models with 7-13 billion parameters on university servers, I measured the power consumption via nvidia-smi and CodeCarbon and then estimated the amount of water used in the data center to power the system, based on publicly available metrics for WUE and grid water intensity in data centers. Afterward, I experimented with various realistic ways to minimize the energy and water consumption, including downsampling the model size, batching input prompts, and setting maximum sequence lengths, and found that slight modifications could result in a reduction of the estimated energy and water consumption per prompt by around 20-30 percent without affecting the quality of results.
But a precise measurement by itself will not solve the issue when the data released to the public has been stripped of all context. This is where the human and social factors become important; the way in which we frame the environmental impact of AI technology affects what people feel they must do, or even be able to ask for.
My STS research revolves around how disclosure practices and metric selection can contribute to an inaccurate representation of the environmental impacts of AI technology. Using the term "misinformation by omission" from Luccioni et al., I conducted qualitative discourse analysis on technical publications, corporate sustainability reports, and media outlets to demonstrate how an in-depth academic analysis, that incorporates both the cooling of the site and the water use at the power plants that provide energy for queries, gets misrepresented in headlines as a "bottle of water per query." Using Langdon Winner's idea that artifacts have politics, I demonstrate how metrics such as the WUE do not remain neutral but are political artifacts that refocus discussion from the appropriateness to optimization of resource use.
In combination, these two projects show something that neither alone could: there are not two separate problems - technological and political - around AI’s resource consumption and the ways it is talked about, but rather only one, which causes both of these problems. Better measurement standards and "minimum viable transparency" requirements, demanding facility-level, scope-specific, and locally contextualized disclosures, are not just good engineering practice. They are an absolute requirement for having any meaningful democratic discussion about the future of this technology.
