AI gedachten / AI thoughts

Last week, I read an interview in a technology magazine, where someone said “I heard that sending one photograph through Whatsapp equals X kilometers in a car”. That triggered my internal factchecker. Turns out that sending one Whatsapp photograph leads to the same amount of carbon exhaust as 0.3 millimeters (!) in a fuel car. Some 15 years ago, similar things were said about Google searches.

The energy usage of AI is much debated. Writing “please” or “thank you” in a chat app allegedly costs around 0,25 Wh electrical energy. Converted to carbon exhaust (0.1 grams), this compares to 70 centimeters in your fuel car – closing the gap in with the car in front of you a traffic jam. Not very impressive.

ChatGPT as steam engine

So, it’s not a problem? Well, no. We need to go back to 19^th century Britain, meet William Stanley Jevons.

Back then, coal was the main source of energy. The steam engine caused a lot of demand for coal and people worried that the coal supply would deplete. Some people suggested that increasing efficiencies, both in mining the coal and in using it (newere steam engines) could solve that problem.Jevons, being an economist, looked into this with a bit more insight. He concluded that higher efficiency caused a disproportionate additional demand. Net effect: more coal would be used. His “Jevons Paradox” is also called the rebound effect, for obvious reasons.

Just like now, a wide variety of applications for steam engines and other motors were invented. Some of these, like moving walkways, are now perceived as quite excentric. It raises the question: how much need for moving things can humanity have? You can’t put a motor in everything, right?

From a motor driven by explosions to an explosion of everything

Well, look around in the room where you are now. In my working room, I count some 10 electrical motors: from the fan in my PC to the clock on the wall. When I take my whole house into account, I easily find more than 60 electrical motors. These motors don’t make noise, they are very efficient and you don’t see the exhaust: the power plant or the battery factory are located far away. That’s not a problem. Or is it?

The rebound effect is really a problem. Especially when the visible costs (noise, exhaust, money) are hidden, there is no limit on the use of a new technology. When computers became abundant in the 80s, storing an extra piece of digital information became almost free: the start of the data-explosion. When internet was introduced in the 90s, distributing information became almost free: the internet traffic exploded. Now we have Generative AI: the creation of information becomes almost free. So what do you think will happen? (I will keep ‘AI slop’ for a future blog.)

Talking about electric motors. Think of all the devices in your home that are connected to the internet. Today, we are more critical about this, not because of energy but because of hackers, but it is still an amazing amount. (Personally, I’m not that advanced about these things, but I still count 15 online devices in my home.)

The digital world mimics the physical world

One question to ChatGPT equals ten Google queries (the classical ones, without the AI summary). Using my favourite metaphor of the car, you could drive 50 meters in your fuel car and have the same carbon footprint as one ChatGPT question: 7 grams CO₂. Well… When I’m feeling lazy, and use my car to drive to the supermarket instead of my bike, I already cause the equivalent of 60 ChatGPT questions. That’s one hour of typing at your laptop. The average citizen will think of one full hour of ChatGPT as slightly overdone (“how much energy does that cost?!”) but 4 kilometers in a car is a very everyday thing.

It gets stranger. When you use that hour to write an article in the classic way, your computer uses more energy than an AI system that writes such an article in a few minutes. AI is more efficient than humans who carry out the same task. It’s better to have AI do the work, and close your laptop the rest of that hour. Or is it?

The real problem

You can feel it coming. We do not close the laptop, we will use the rest of that hour to continue working. We do extra things. And those extra things will probably involve AI as well. (Hi, Jevons! Nice to meet you again!)

The patterns in our behaviour change – it’s called a ‘systemic change’. That car that I just presented, it does not reduce the time we spend on our daily commutes. The travel time remains the same, it’s just the distance that increases. Energy use increases, too: transport accounts for 30% of the EU energy use.

Wise use of AI should avoid that AI is going into the same direction. The real impact on energy use happens when we put AI in everything around us.

Wise use of AI is possible: improve healthcare, optimize transport, information that is easier to understand. AI can be used to reduce our energy need, but let’s not forget to listen to Jevons from time to time.

(NB The numbers in this article will be outdated soon: AI computers and cars both become more efficient. That’s good – but the rebound effect will always be there!)