We Don't Even Know How Bad It Is Yet

Before we can talk about solutions, there is a frustrating reality to address first. We genuinely can't fully measure AI's environmental footprint yet.

Researchers Luccioni et al. (2023) studied the BLOOM language model and found that training-only carbon estimates, which have been the standard, captured less than half of the model's actual carbon cost. Even more surprising, about 75% of energy during live deployment went toward just keeping the model running in memory, not answering your questions. The server is consuming energy whether you're using it or not.

The MIT Climate and Sustainability Consortium put it plainly: AI is developing faster than researchers can measure its tradeoffs. This isn't a political problem or a funding problem. It's a knowledge gap, and it's a big one. You can't regulate what you can't quantify, and you can't quantify what companies won't disclose.