The environmental impact of AI is increasing rapidly – its disadvantages receive too little attention
The use of AI has quickly become a part of people’s everyday lives, and you, too, probably use it either at work or in your leisure time. Do you know, however, how much the use of AI burdens the environment and produces emissions?
This question is so difficult that not even University Teacher Veikko Halttunen, who teaches responsible usage of information systems at the University of Jyväskylä, knows the exact answer.
It is nevertheless clear that we are dealing with vast numbers. The energy consumption of AI can even be compared to that of middle-sized countries.
In his work, Halttunen has examined the environmental impacts and responsibility of digitalisation, and in his teaching he has preferred not to emphasise the environmental impact of any particular technology, but when it comes to AI, he has had to make an exception.
“I have sought to deal with the responsibility of digital systems at large, but AI, as a particular technology, has become such a major consumer of energy that it cannot be ignored,” Halttunen explains.
The disadvantages of digitalisation must be recognised
Halttunen says that he wants to make people aware of the various drawbacks of digital systems as well. One of these drawbacks pertains to the environmental impact of AI.
We need to understand and recognise the negative aspects in order to address them.
“We have developed a bias towards ignoring the drawbacks of digitalisation and digital systems. Digitalisation has taken society forward and created employment. But in spite of this, it has negative sides as well,” Halttunen notes.
AI is not a chat tool but a computational system
When talking about the environmental impacts of AI, Halttunen first wants to dispel the common misconception that the training of an AI model consumes more energy than its use:
While training is itself highly resource-intensive, the volumes of usage are so huge that they quickly exceed the load caused by training."
An AI model refers to a mathematical system trained on large volumes of data, in this case, a generative language model. An example of a popular AI language model is ChatGPT, developed by OpenAI.
Halttunen says that the energy consumption and emissions resulting from the use of AI can very quickly reach a ten-fold level compared to the energy needed for training, likely within just a few months.
“For many people, it seems to come as a surprise that the use of AI can consume so much, because it is such an abstract and immaterial thing. The present AI systems require vast computing capacity, however, and their usage volumes are vast,” Halttunen says.
Halttunen wants people to understand that AI is not a chat tool, but a computational system.
Creating images and videos consumes a lot of energy
In the context of AI use, we talk about prompts, that is, instructions or questions given to a generative language model. An individual prompt does not consume much energy in itself, but the widespread use of prompting has significantly increased the overall energy consumption associated with this technology.
Because of the energy demands of AI models, large numbers of data centres are needed in Finland and across the world.
Halttunen also points out that the energy consumption depends greatly on the scope of the AI model and what it is used for.
Generating images and videos consumes a lot of energy.
It has been estimated that nowadays tens of billions of images are generated using AI each year. As for videos, fewer estimates are available, but they are also produced on the scale of tens or hundreds of millions per day. In terms of energy consumption, one video may equal to generating up to hundreds of images.
AI developers withhold energy consumption data
The exact energy consumption of AI models is a mystery to researchers as well, as it is difficult to obtain any exact figures about overall consumption.
OpenAI, which has developed one of the most popular AI models, estimates that generating a simple query by means of their ChatGPT model consumes about 0.34 watt hours of energy. It would be roughly comparable to a Google search.
Some researchers have estimated, however, that a ChatGPT query could consume up to ten times more energy than a Google search. This estimate was made by researcher Alex de Vries in 2023.
Halttunen criticises AI operators for a culture of withholding information as they do not disclose their energy consumption figures. In addition to the model developers, this also concerns the whole ecosystem, including data centres, for example.
“The problem for an ordinary consumer is that these issues are not widely discussed. As a result, people may get the impression that the use of AI does not involve any significant environmental impact.”
The issue also affects researchers, who have to piece together information from limited sources and draw on data that is several years old. This also leads to researchers having to make too many assumptions when conducting their research.
“To serve the best interest of the humankind, the data should be easily available to researchers – and also to ordinary citizens”, Halttunen summarises.
AI is a more energy-intensive technology than block chains
According to Halttunen, one way to illustrate the scale of AI’s energy consumption is to compare it with another widely discussed technology, block chains, which are particularly associated with cryptocurrencies.
“In contrast to the use of AI, block chains are a very marginal technology. Yet, mining a single cryptocurrency, Bitcoin, currently consumes approximately 180 TWh per year. This is roughly on par with the total electricity consumption of Poland or Argentina,” Halttunen states.
According to Statistics Finland, the annual electric consumption in Finland is about 83 TWh, which means that Bitcoin mining consumes more than twice as much electricity as the entire country of Finland.
AI is a considerably more energy-intensive technology than block chains. Its usage is also increasing rapidly, and people generate more and more images and videos using AI.
Emissions from the ICT sector will grow significantly in the coming years
The climate emissions of the whole ICT sector are already comparable to those of air traffic, while its energy consumption is clearly higher.
It can be expected that the AI boom will rapidly increase the ICT sector’s share of global emissions.
According to various forecasts, the energy consumption of data centres, for example, is expected to more than double by 2030 because of AI.
In addition to the energy required to train and use AI models, it is important to recognise that AI also contributes to climate emissions in other ways, such as through land use and the cooling requirements of data centres.
The researcher’s three tips for considering the environmental impact of AI in everyday use
Despite its negative effects, Halttunen does not advocate banning the use of AI and notes that he also uses it himself in his own work.
However, he highlights three key pieces of advice, plus one additional tip, that everybody should follow if they wish to take the environmental impacts into account in their own AI use.
- Use AI only where needed. Don’t generate unnecessary content.
- Avoid generating images and videos, since this clearly has the greatest environmental impact.
- Plan the prompts beforehand and think carefully what you want to ask. Try to ask as few questions as possible.
Additional tip: Avoid using AI over a mobile network, favour fixed connections instead. Data transfer via mobile networks consumes much more electricity than through fixed networks.
Halttunen points out that paying attention to the negative aspects of technology also helps take them better into account in society and in the development of different technologies.
“It is important that everybody understands how AI works and why it has such a big environmental impact. Understanding AI helps us make more environmentally-aware choices regarding its use,” Halttunen concludes.
