- One everyday prompt uses about as much electricity as a microwave does in a second.
- A typical user's 4 prompts a day take under 0.02% of the electricity a UK home uses in a day (A heavy user at 30 prompts per day is more like 0.3%)
- Boiling the water for a mug of tea is the equivalent of 130 prompts.
- One electric shower is worth about 3,500 prompts.
- On-site cooling water is about six drops a prompt.
How does that make you feel? Is that a lot? Is it a little? Does it change how you think about how you use AI?
After every keynote I give now, the first question is always about the environmental impact of AI. Sometimes it's thoughtful. Sometimes it arrives as a note tied to a metaphorical brick, thrown at me from the audience as a protest at what some people misconstrue as my blind enthusiasm for AI in a modern society.
I don't think the question is wrong and I actually share some of the concerns. But what I worry about more is how poorly equipped we are to answer it.
Few of us, me included until recently, can say whether the statement "AI is unsustainable" is true, or even what numbers sit behind it. And frankly, given that the questions are not going to go away, I decided to see if I could find out.
The result? A whitepaper, available here, that attempts, in plain-English, to take an open book approach to creating some benchmarks that we can all relate to, from which we can make our own decisions about how much it impacts the environment, and whether we should modify our behaviour.
Of course, it's complicated, not everyone publishes the data we need to come up with an answer with a narrower range or higher degree of confidence. Only one provider has measured its figure across a real production service. The rest are modelled or unexplained, and none is independently audited. And what is a prompt anyway? Some of my prompts are a sentence, others are a 10 page document. Heavier tasks cost far more: a reasoning prompt is about 13 times an everyday one, and a short video can be thousands of times more. And what about training the models? And the water used for cooling not to mentioned the water used to generate the electricity, these are the least certain figures of all.
Then there's scale, because when a small number gets multiplied by a really, really big number, you end up with a big number. My estimate is 4 to 8 billion chatbot prompts a day. That's the daily electricity of somewhere between 175,000 and 350,000 UK homes, and more than twice that on a more cautious model. Worldwide, data-centre demand is projected to roughly double by 2030. Each task is getting dramatically more efficient, and total demand is still rising. Both are true. None of those totals tells you what it's like to live next to a new data centre, where the worry is local (the water, the grid) rather than global. That's a different debate from the one this paper is about.
But, I think there is an approach, that helps us understand, maybe not an explicit number, but a guide, a benchmark we can use to be better informed about the actual impact of our AI habit.
And we can learn from other industries, we can learn from the automotive industry about the difference between the impact from making the thing, vs driving the thing. Tesla publishes this split and we can learn from it. They disclose that the impact per mile of driving a Model 3 or Y, is about 132g of carbon (Tesla's own figures, for the Europe Long Range version, assuming the grid stays as it is today). 94g of that total comes from the manufacture of the vehicle, only 38g comes from the electricity you charged it with. (Just for comparison, it's 260g for an average petrol car, and that's just for the mileage it doesn't include the impact of making it).
The same is true of AI. Training a model is the making, your prompts are the driving, and every number at the top of this piece is the driving. Spread over a trillion prompts, training and hardware add roughly a third on top. Spread over 100 billion, they could add more than the driving itself, and for the newest, largest models it could be more again. Nobody publishes how many prompts a model serves in its life, so it's the weakest number in the paper.
We can learn from the food industry - which has the lower carbon footprint for the same amount of protein: salmon, chicken, pork or beef? Most of us couldn't say. On global averages, farmed fish and chicken come out at roughly 6kg of CO2e per 100g of protein, pork at about 7.6 and beef from beef herds at around 50. So beef carries roughly 8 to 9 times the carbon footprint of chicken. And yet, how many of us understand that never mind make changes to our eating habits as a result?
Part of my point is that some of the highest-impact choices we make about our own individual impact on the planet are among the most ordinary, and we never get shown the numbers. That's the gap I wanted to fill. Yes, there are assumptions, and yes, the range is huge - 4-8 billion prompts a day is one hell of a swing. But it's a start.
You know that I'm clear about my own position: I'm AI positive but I am also a big fan of life on earth and like everyone, I'd quite like it to continue if that's OK. I'm not interested in declaring whether AI is good or bad for the planet, and if you look, you'll see the paper doesn't either. The arrival of AI in our lives (as with every other major technology since the taming of fire) is a judgement for society, let's at least make it by being informed.