AI & the climate

Mr Creutzig, in your definition, what exactly does climate protection entail?
Essentially, climate protection has two dimensions – mitigation and adaptation. Mitigation means the reduction of CO2 and greenhouse gas emissions. In the context of adaptation, we take measures that mitigate the effects of global warming. Adaptation has its limits, e.g., we can raise dikes, but if sea levels rise too high and storm surges are too severe, at some point, this will be of no use. Similarly, we can only protect ourselves to a limited extent against heat waves.

In this sense, mitigation is more important than adaptation: it is better to prevent such severe impacts from occurring in the first place. And with that in mind, climate protection means not only stabilizing greenhouse gas emissions or a certain reduction, but actually net zero emissions.

When did the idea first arise that AI could help in this area?
People tend to associate all modern technologies with climate protection. This also applies to AI as a "general purpose technology" – a technology that can be used anywhere. However, the overall picture is far more ambivalent.

An exciting moment occurred in 2019 when a team from Silicon Valley published a study on the topic. Coders with academic backgrounds had formed a movement that wanted to pay increased attention to where and how AI could be used for good – for climate protection, for example. These first joint efforts on systematization are particularly interesting because climate protection is not what primarily drives the Silicon Valley community.

Can you give a concrete example of how AI can promote climate protection or help us become more climate-friendly?
My team and I are currently working on the use of AI for sustainable urban planning.

Until now, climate change scenarios have always been designed on a national or global level which is not very helpful for cities and their administrations. However, thanks to the now countless data that tells us something about how cities function at street and building level, it is possible to calculate exactly how cities can be made more climate-friendly through urban planning.

Here AI contributes to climate protection through data processing and evaluation. What new insights have you gained in this way?
Conceptually, we already knew a lot about sustainable urban planning. For example, that we should densify along local transport axes, such as underground railway lines, i.e., use vacant areas within existing buildings. Thanks to AI, however, accuracy has increased. It can predict faster and more accurately where new housing will produce the least greenhouse gas emissions. Thus, they can be allocated to optimal locations.

Have you already used AI to work out recommendations for action for a specific city?
Indeed, we were able to prove for Berlin that so-called sub-centers play an important role when it comes to reducing VKT (vehicle kilometers travelled) – i.e., the number of kilometers travelled by car.

In our study, we developed an explainable machine learning algorithm that analyzed the data of 3.5 million car commuter trips as well as high-resolution urban design data. The result was that the so-called "15-minute city hypothesis" is true: The idea is to (re)design cities in a way that allows residents in all parts of the city (not only in the city center, but also in the suburbs) to meet their daily needs – i.e., shopping, working, going to the doctor or to school – within a 15-minute walking or cycling radius. The more this is given, the less they use the car.

So, AI can help us to plan and act in a more climate-friendly way. But to what extent can AI also harm the climate?
AI potentially has two effects: On the one hand, efficiency gains – you can do things faster or achieve the same result with fewer resources. This tends to be climate-friendly because resource efficiency usually also means CO2 efficiency.

On the other hand, we have the scaling effect – if something can be done faster and more efficiently, people do more of it, and this can cancel out or even overtake the first effect. Furthermore, AI promotes general consumption via algorithms, e.g., on social media, and thus also increases the consumption of resources.

Does this mean that the harmfulness or friendliness of an AI to the climate is primarily a question of its application or use?
Exactly. There are areas where the use of AI is unquestionably good for the climate – for example, as just mentioned, in sustainable urban planning. Or in agriculture, where precise mapping of soil quality helps us know where to use fertilizer and how much.

Other applications, however, are far more ambivalent – autonomous driving, for example. Autonomous driving is highly AI-reliant but has in the first instance no sustainability benefits. On the contrary, autonomous driving has the effect of eliminating the driver's time costs. Thus, without regulation, the danger is that even more people will drive.

How could the resulting damage to the climate be reduced?
For example, by only allowing autonomous vehicles for certain purposes, such as shared taxis.

After all, a car is a terribly inefficient machine – It weighs about 2 tons and transports an average of 1.7 people, or 80 to 90 kilograms, with an engine that is currently 20% efficient. That will improve somewhat with electromobility, but still: if I take my bicycle, I have 14 kilos of material to drive around with, which compared to a car is better by a factor of 100.

You can improve this inefficiency with shared taxis (3 to 4 people in one car). In order for this to catch on, however, we need regulations, especially in urban areas, that prohibit the use of a car by a single person, for example. Or that forbid the provision of public parking spaces. That would be important, but not everyone likes to hear it.

In your example, regulation would again concern the application rather than the AI behind it. Couldn't we make the AI itself more climate-friendly? E.g., by defining that it must have a certain energy efficiency in order to be allowed to be used at all?
Of course, you can also calculate how energy-intensive an algorithm is and think about how you can make it more efficient. For example, instead of training an algorithm over and over again, you can update it, i.e., only use the novelty value of additional data. However, the influence of such measures on the resource efficiency of an AI is relatively small compared to that of the application.

It would therefore be valuable if software engineers were more concerned with the regulatory conditions of the application and tried to think in this context. For example, what new usage patterns emerge, what this means for energy consumption, and how sharply rising energy consumption can be avoided. In turn, the regulators also need to understand what AI possibilities exist. The exchange here definitely needs to be improved.

In terms of AI ethics, new regulations are underway – e.g., the EU AI Act. But this is mainly about ethical issues, right? Or also climate protection?
As far as I know, climate protection has not yet been properly considered in the EU AI Act, so there is a need to catch up. As I said, however, I believe that this is not only a matter for the EU, but also for individual cities and municipalities.

What I would like to emphasize in this context is that ethics should never be an afterthought but should already be considered during the development of an AI. I believe that is the challenge: that ethics, and with it climate protection, is not an additional requirement, but that the AI developers themselves are made aware of it – ideally already in their training.

All in all: Do you think AI is more of a “climate helper” or a “climate sinner”?
I think it is very important to pay attention to the governance of AI. And to the possibilities of all stakeholders – policy makers, businesses, and civil society – to institutionalize the handling of AI. This is where we can act and where we need more capacity and awareness.

To conclude: What important factors do you find are often overlooked in this topic?
That here, too, as we show in our latest study, it's about the bigger picture that at first glance has little to do with climate protection.

Like social inequality and polarization: AI plays a central role in both – we know that societies are becoming more unequal through AI and digitalization in general. The other is that polarization effects are often AI-driven, especially in social media.

At first glance, this has nothing to do with climate protection, but then again it does: for more climate protection, we need a social consensus, a political discussion – everyone must and should actively participate in this difficult change. But if increasing divisions makes people lose social trust, it is very difficult to achieve this. In this context, the question of the impact of AI and digitalization on society as a whole is also important for climate protection.