Big tech’s selective disclosure masks AI’s real climate impact
This seems spot on:
Using any sort of statistical summary of the data, rather than the aggregated energy and climate impact across the whole system, will always give a misleading view. They mention their data is skewed, but they don’t mention in which direction. If there is a material number of high-energy ‘reasoning’ prompts skewing their dataset, that means the total energy consumption of all prompts will be very high, with much of the responsibility coming from a few energy-hungry queries.
Part of the reason this is important is that this week, we saw a new research paper that shows that the energy consumption of text generation massively increases for every small gain in accuracy from the use of energy-hungry ‘reasoning’ models:
It would have been pretty easy to supply the range, the skew, the average and the median, or even the actual entire dataset, to avoid any doubt. Any hint of looking at the broader system rather than individual responsibility is excised from this paper. That is clearly an intentional choice: if Google disclosed the system impacts of generation, it would probably look way worse. [....]
The per-query narrative framing paints the precise opposite picture to what we see when we look at what really matters for environment and climate: the absolute figures.
Regions with high data centre concentration are seeing accelerated growth in power demand that incentivises fossil fuels, either slowing down climate progress or reversing it entirely. The sphere of that influence is expanding from towns, to states, to countries. The companies that own them can only partially hide the steep backsliding in their aggregate disclosures.
Renewable energy that should be displacing fossil fuels ends up meeting new data centre demand, granting coal and gas extra years and decades of immediate, measurable harm to human life. The worst players don’t even bother with the grid, plugging data centres directly into new, custom-built fossil fuelled power stations that’ll hurt people for decades after the hype dissipates.
Tags: llms ai environment energy climate climate-change google meta amazon openai datacenters