IS USING Ai LIKE DRIVING A GAS GUZZLER?

We poked a digital hornets nest.

For me, that was the core question behind a recent Rare Earth. We wanted to find out if our own digital behaviour and especially the growing use of artificial intelligence is the new climate scourge. The energy demand for existing and, especially, new data centres is eye-watering, and unlike getting behind the wheel of a chunky 4x4, it seems to be elusive, dodging criticism and scrutiny. Also on the digital domain’s environmental charge sheet are excessive water use and habitat destruction, both explored in the show (link above), but here I’ll focus on the wattage.

It was a lively programme and I came away pretty convinced that Ai today is a bit like oil in the early 20th century - a recently discovered power on the brink of an economic coup. But this time we know the harmful side effects, so should strive to limit them.

Audience reaction was very positive but there was sharp criticism from some in the computer science world saying we had got our figures wrong and misunderstood the basics of data centre energy usage. So, before we start the next Rare Earth series in mid-August, I thought I’d dive in.

(Just before getting into the argument, I should declare a bias: I don’t like Ai. I like human imagination, teamwork and creativity. And I like Ai even less when I can’t tell the difference. It’s also trained on the couple of books I’ve written without permission or reward.)

Shortly after broadcast, Dr Daniel Schien, an associate professor at the School of Computer Science at the University of Bristol, contacted my co-presenter Helen Czerski to say this was an important topic but one of our experts provided ‘misleading and factually incorrect information’ He went on to write a post on LinkedIn Public Disinformation on Ai Energy in Rare Earth. You get the drift. One of the statements in our programme he disputes is that data centre demand could be on course to exceed all electricity supply (yes all) by 2033. In my defence, I did react on air with some incredulity to this claim. I gave Dr Schien a call.

“This [claim above] is based on a misunderstanding of data centre energy consumption. In the context of reaching our climate goals, the growth of data centres is a real problem…but some people are spreading false information...and sensationalist messaging is poisonous”

He says that these enormous future figures come from looking at energy demand and data flows in 2011, then looking a data volume projections and multiplying the 14 year old energy number by the growth forecast.

“But data volume is not a reliable proxy for energy use” he tells me. He explains that so much of the energy demand associated with energy centres is from their construction and designed capacity - in effect their existence rather than their use. To help out let’s try a couple of analogies:

1. Brains. It takes a lot of energy to create it and maintain it. But using it intensively by thinking hard - processing large volumes of data with greater neural activity - only causes marginal increases in calorie demand.

2. Trains. Nearly all of the energy demand from the rail network comes from building the infrastrucure and moving the rolling stock. A busy train, packed with passengers doesn’t add much to the electricity or diesel demand over an empty one. Your choice to take the train has a vanishingly small energy consequence.

This analysis leads Dr Schien to say we shouldn’t beat ourselves up about sending emails with large attachments or watching videos in 4K as “We can barely affect the energy consumption of networks”. Though he does remark that using Ai to generate videos or graphics does immediately demand considerable wattage.

But I’m not sure this lets us off the hook. Our behaviours today drive the building of new infrastructure tomorrow. It might be responding to a viral post, storing thousands of photos in the cloud or using Ai to re-imagine my children as kittens. As more of us behave this way (often urged on by those very tech companies who make money from our digital cravings) more data centres are needed to handle the processing. Back in the late 90’s I was BBC News Transport Correspondent and the mantra for new road building then was ‘Predict and Provide’. This meant ‘Lets look at growth projections and lay more miles of tarmac accordingly’. It turns out that having more roads encourages more driving.

Let’s reverse the thought. If we didn’t use the services, they wouldn’t build the providers. Just as declining train passengers would mean no new railway infrastructure.

So, individual actions are trivial but a collective global boycott of frivolous Ai could bend the curve away from climate threatening energy demand. I won’t be betting my house on it.

Meanwhile, like so many things in fighting climate change, it’s the appetite of government and business for Ai which will have by far the greatest impact. And, judging by our PM’s swooning over the technology and feting the tech companies, they are betting our future on it.

Where all sides in this debate comes together is the need for total transparency from big tech on the energy usage of the all growing components of the digital world. This would enable all of us to shape our behaviour based on facts. Less speculation more information.

ps There has just been a good accessible UNESCO paper on this very subject.