I’m trying to make sense of some of the figures shared when we talk about the environmental impact of AI, and what the specific drivers of it are, in terms of the specific kinds of AI being deployed around the world. Here are some notes, and a request for help to point somekind of useful taxonomy for thinking about the energy requirements of different ‘flavours’ of AI.
In the book the AI con, instead of talking about AI, Emily Bender and Alex Hannah talk about what activity is being automated as a way to have a more precise and productive discussion about what AI is used, how we end up prioritising public support for it ahead of other things.
I wondered if this might be helpful for having a productive discussion, in the context of some eye-opening insights from a recent report
Here’s an excerpt from the book that I think is helpful where they use a taxonomy of things being automated, instead of a vague catch all term of AI:
- Decision making. The first group involves using computers to automate consequential decisions. These are called automatic decision systems and they are often used, for example, in the process of setting bail, approving loans, screening résumés, or allocating social benefits. These uses are contentious, and rightfully so, because they have extreme ramifications for people who are subject to the system’s recommendations.
- Classification. The second kind of automation involves classification of inputs of different types. For example, image classification can be used to help consumers organize their photos (where are all the photos of Grandma?), or can be used by governments for surveillance (matching a security footage frame to a database of driver’s license photos). The classification of web users for targeted advertising also fits into this group.
- Recommendation. A third type selects information to present to someone, based on their own search or purchase history, or searches performed by someone else with a similar profile to them. These systems are called recommender systems. They re behind the ordering of your feed in social media websites, Amazon product recommendations, or movie suggestions on Netflix.
- Transcription/Translation. The fourth type is the automatic translation of information from one format to another: automatic transcription (sometimes called “automatic speech recognition” or “speech to text”), finding words and characters in images (like automatically reading license plates), machine translation of one language to another, or something like image style transfer (taking a selfie and making it look like an anime character).
- Text and Image Generation. Then finally there’s a type that’s been very much in everyone’s mind recently: so-called generative Al or, more aptly, synthetic media machines. These are systems like ChatGPT, Gemini, or DALL-E that allow users to generate images or plausible-sounding text based on textual prompts. A “prompt”, in generative Al terminology, is the words used to describe the desired output.
If we refer back to the that wild GenAI vs Trad AI chart from the Schneider Electric report, we see a clear divergence between in resource footprint between what you might assume are the first four, where energy usage is assumed to be relativelt flat, and the new fifth one, which if you follow the Schneider report, is responsible for almost all of the growth. For convenience, I’ve added the chart below, from the most optimistic, Sustainable AI scenario.
What is this chart showing us?
The bright green line shows us the total, the orange and yellow show the figures for Gen AI inferencing and training respectively – add them up and you have almost all the energy use there’s a tiny token amount for the “trad AI” training and inference by comparison.
Is this fifth ‘generative’ category the problem category in terms of direct environmental impact?
This seems somewhat overly simplistic, but at the same time, in the last post I wrote about had two sources, David Rolnick and the Schneider Electric report essentially saying something similar. I’d love to see some more helpful nuance out there.
Is this the only problem to think about?
Of course not – in the “taxonomy of automation” above, with first 4 groups of activity have plenty of scope for harm, beyond the direct environmental impact consequences. In fact this taxonomy of AI, algorithmic, and automation harms below lists plenty of things to be mindful of. See this published paper listing them all, or see the database tracking news stories sorted by this taxonomy.
However, when talking about the direct environmental impacts from meeting the astonishing amounts of electricity demand, you do get the impression that Gen AI is the key driver, representing a qualitative shift in both the direct energy demand, but also in the visibility of AI as a thing that a growing number of people use regularly.
Is this just down to GenAI being the UI we use to consume AI services?
One thing I don’t understand though is to what extent Gen AI interfaces end up being the interface we see as end-users, that causes us to over-index on that specific use of the technology, versus something massively more widespread, like ranking and recommending that we see used daily in search and e-commerce systems.
Before I saw the Schneider Electric chart, I could easily believe that most energy consumption from AI might be from recommender systems for example, but without any access to meaningful numbers in the public domain, it’s basically impossible to know.
A call for help
If you have come across a taxonomy that helps us reason about the relative power consumption of different ‘kinds’ of AI, and the share they make up of wider commercial use, I’d love to hear about it.
It’s a huge gap in terms of the data required for a meaningful data-informed discussion about AI.