The environmental impact of AI, the debunking of the environmental impact of AI and the debunking of the debunking
At the recent Deep Learning IndabaX South Africa, I was sitting towards the back of the auditorium when one of the speakers, a researcher from IBM, made a comment about the staggering amount of potable water used when serving LLMs like ChatGPT. I can't even remember which of the commonly cited figures the speaker used, but it was probably the one that says ChatGPT consumes, on average, half a litre of water per query. Granted, my attention to detail when listening can be lacking, but I think the reason that I care so little for the exact estimate is because the metric itself is flawed.
When someone says something that may reveal an uncomfortable truth, we naturally want that to not be a truth after all. We have to debunk it! And so if you Google 'chatgpt water use per query', you’ll find a lot of debunking. From my vantage point at the back, I could see a lot of shaking heads because, can you believe it, the speaker just repeated a debunked claim and everybody knows that! The guy sitting in front of me actually used ChatGPT to fact-check the claim in real time and was visibly thrilled when it obliged with a thorough debunking! (Using an LLM for a fact check - at a deep learning conference no less - seems problematic, but I digress.)
The points made in the debunkings that I've read can be summarised as follows:
Sam Altman himself recently entered the debate and provided his own estimate of ChatGPT's water use in an evidence-free blog post: 0.000085 gallons, or 0.3ml, per query. The message is clear - water usage is negligible, don't worry about it.
The problem with the original claim and the 'debunking' thereof is that the denominator is wrong. The costs of training and serving AI - that is, the amount energy needed to power data centres and the fresh water required to cool them - are not shared equally by the users. Instead, the costs are highly concentrated and borne by those communities where tech companies have built (and are building) mega data centres. As Karen Hoa details in Empire of AI, some of the biggest data centres in the world are located in some of the most water scarce places on the planet. The 'per query' or 'per user' estimates don't say anything about the impact on the (often already vulnerable) communities who bear the costs.
Another problem with the "compared to this or that other industry, AI is not so bad" argument is that it completely ignores the fact that this technology, and thus its environmental impact, is new, a marginal cost. We therefore should be asking whether the additional cost is justifiable.
As someone in this field, someone who believes in the potential of technology to help people, I understand the urge to want to push back when it seems that the negative impact of AI is overstated compared to other industries. But by succumbing to our confirmation bias so easily and eagerly accepting the 'debunkings' without any critical thought, we effectively say AI has no environmental impact, that it is not something to be concerned about. So while some individual estimates may be wrong or exaggerated, the environmental impact is real, and as with everything related to the climate crisis, the costs are not shared equally.
By explaining away the environmental impact and satisfying ourselves that it is not something to be concerned about, we are not sufficiently incentivising fundamental research into new AI paradigms or applied research and product development into small, efficient, and distributed models aimed at solving real problems. Even if the 'prophets' at OpenAI and Anthropic say the scale is necessary to cure cancer, I certainly don't need it to debug my SQL.