Freiburg, 08.08.2024. Artificial intelligence (AI) has been on everyone’s lips since the release of the first version of ChatGPT at the end of 2022. Intelligent software applications have already been in use for many years: compiling favourite playlists, sorting out spam emails, finding the best route, text translation or product recommendations based on previous online purchases. But what about the carbon footprint of AI?

Jens Gröger, Senior Researcher in the Products & Material Flows department, explains the advantages and disadvantages of AI when it comes to environmental benefits in the Oeko-Institut’s latest podcast ‘Wenden bitte!’. To the Öko-Institut podcast ‘How sustainable is artificial intelligence?’ [https://www.oeko.de/podcast/wie-nachhaltig-ist-kuenstliche-intelligenz/]

More efficient processes versus increased energy requirements

In contrast to traditional computing, machine learning is based on very large amounts of data and parallel computing processes. This goes hand in hand with increased computing power. Currently, around 1.5 percent of Germany’s electricity requirements are used for data centres alone. This demand will continue to rise in the future. This is because computer applications are currently being equipped with more and more AI functions. For example, a query via ChatGPT consumes three times as much power as a traditional search query. If AI functions also find their way into normal office applications, such as text and image editing programmes, their power requirements will increase considerably. The environmental impacts occur both during the training and operation of AI systems. The training of ChatGPT in version 3 alone is estimated to have caused 500 tonnes of CO2, with a single request accounting for around 4.5 grams of CO2.

On the other hand, AI has the potential to optimise technical processes – such as the production, maintenance, use and ultimately waste sorting and reuse of products – to help save energy and resources and promote the circular economy. AI can also help to optimise the use of wind and solar energy in the energy sector. But do the positive effects outweigh the disadvantages? According to the scientist, these questions are still unanswered in many cases and require further research on the one hand and legal regulation on the other.

Measuring the environmental impact of AI

In order to measure the carbon footprint of AI, the scientists distinguish between three levels:

1. Direct effects that can be directly attributed to digital technology. These include the production and use of end devices, data lines and data centres.
2. Indirect effects that are related to the use of digital applications or AI: In online shopping, for example, these include packaging and delivery; in the optimisation of production processes, it is the reduced energy requirements.
3. Systemic effects that affect society as a whole, such as changes in mobility behaviour due to car-sharing services or changes in the world of work. These also include rebound effects, i.e. savings in one place that are accompanied by increased consumption elsewhere.

The environmental impact of direct effects is best calculated. Indirect effects can be estimated on the basis of use cases. However, systemic effects have so far been difficult to quantify. Jens Gröger is in favour of a life cycle assessment approach: ‘In a life cycle assessment, we examine the entire life cycle of a product, from raw material extraction and production to transport, use and disposal. This methodology can also be applied to digital applications such as software and AI.’

Transparency as a basis

With the knowledge of the environmental impact of digital applications, the second step is to reduce energy consumption. ‘When it comes to digital technology and AI, we can’t just let technical development run its course,’ says Jens Gröger. ‘This can vehemently go in the wrong direction. A technology impact assessment and regulation based on this are essential. Undesirable developments should be recognised at an early stage before they become uncontrollable.’ The scientist is in favour of providing environmental product information with every digital service, for example in the form of a small data package with information on energy and resource consumption as well as greenhouse gas emissions. Users and, above all, reporting companies can then track and evaluate their carbon footprint and other environmental impacts and take appropriate measures to improve their balance sheet.

Knowledge instead of everyday advice

The Oeko-Institut’s podcast ‘Wenden bitte!’ is aimed at anyone with an interest in politics and the environment from the worlds of politics, science, media, NGOs and the general public. The podcast is hosted by Mandy Schoßig, Head of Public Affairs & Communications, and Hannah Oldenburg, Digital Communications & Social Media Officer at the Oeko-Institut. For around an hour, they talk to an expert from the Oeko-Institut about upcoming sustainability transformations – enough time for the ‘long haul of environmental podcasts’. The special episodes address current political and social issues.

More Informations:

https://www.oeko.de/podcast/wie-nachhaltig-ist-kuenstliche-intelligenz/ Episode 5 „Wie nachhaltig ist Künstliche Intelligenz?“ mit Jens Gröger, erschienen am 8. August 2024