Laura Ruotsalainen , Professor of Computer Science at the University of Helsinki, cannot do much to influence how consumers use artificial intelligence, but she can influence how much electricity is consumed in training AI. And that requires mathematics and colleagues.
A single neural network – that is, a mathematical model – can contain millions of parameters, representing different settings on the programme’s command line, and they all influence the end result, for example in pattern recognition.
On closer inspection, some of the parameters may prove to be unnecessary, in which case they are omitted. If a parameter cannot be omitted, its weight can be reduced by optimising other parameters.
"This means that more information can be obtained from a smaller amount of data," explains Ruotsalainen.
This, in turn, reduces the amount of energy required for computation and lightens the environmental impact of artificial intelligence.
“A data centre in Finland is more environmentally friendly than a data centre elsewhere”
The use of AI places a burden on the environment in many ways, some of which are still unknown. At the same time, AI supports the reduction of carbon dioxide emissions and the search for solutions to the ecological crisis.
This dilemma sounds familiar from centuries past: humanity has faced, and seems set to continue facing, the same problem with most of its inventions. Good and evil are packed into the same package, and one cannot simply choose one option over the other.
The dilemma of artificial intelligence – or digitalisation in general – is illustrated, for example, by the fact that measuring environmental impacts requires a computing infrastructure whose own environmental impacts are only partially understood, such as carbon dioxide emissions and water consumption.
“We are not yet able to comprehensively calculate how the physical machines inevitably associated with AI, their components, or the premises required by these machines impact the environment through mining, earthworks and construction,” explains Laura Ruotsalainen.
However, many factors have been successfully measured with precision, which in turn enables, for example, closed-loop energy systems and, consequently, ecological urban planning – a research topic of great importance to Ruotsalainen.
For example, LUMI, the world’s ninth-largest supercomputer located in Kajaani, generates a fifth of the energy required by Kajaani’s residents, as the waste heat produced by the machine is channelled into the homes of the town’s residents. One hundred per cent of the electricity required by LUMI is generated by hydropower, i.e. from renewable sources.
– Individual supercomputers and entire data centres can and must be designed and implemented in such a way that the impact on the environment is kept to a minimum. That is why I dare say that, from a global perspective, a data centre in Finland is a more ecological solution than a data centre located elsewhere, says Ruotsalainen.
– In the case of LUMI, the closed water cycle is 100 per cent. Elsewhere, this hasn’t yet been fully achieved, but in Finland we always strive for it, and in principle we know how to do it. In countries in the global south, they may not necessarily know how or be able to do it, so the cooling water evaporates into the air, which further exacerbates the water shortages in those countries.
Ordinary citizens – that is, us consumers – have become so enthusiastic about AI at work and in our free time that, in just one month, the daily use of AI consumes as much energy as was used to train the entire programme.
In February 2023, the ChatGPT3 programme was used 590 million times; by February 2025, this figure had risen to 1,600 million times. The growth appears to be continuing exponentially.
If only a single question were asked of the AI during each use, or session, it would consume 1,200 megawatt-hours of energy per month. By way of comparison: the average Northern European uses a total of 1.6 megawatt-hours of energy per year for their entire life.
AI programmes can be reused
Making computing more efficient in the way Laura Ruotsalainen proposes requires a deep understanding of mathematics. Fortunately, this expertise can be found within Ruotsalainen’s own research group.
The group consists of data scientists and computer scientists, both of whom work in applied mathematics. All members of the group have studied mathematics extensively.
Laypersons may have come to believe that data science or artificial intelligence somehow replace mathematics, but according to Ruotsalainen, this is not the case.
"Machine learning cannot be developed without a strong grasp of mathematics. That is why we have also been involved in designing tailored mathematics courses for students aiming to become machine learning experts," says Ruotsalainen.
She returns to the topic of the energy consumption of AI programmes.
– We can also reuse AI programmes, meaning we don’t always have to teach the programme everything from scratch, Ruotsalainen continues.
In technical terms, this is known as ‘transfer learning’. When data, machines and energy are available, people are too quick to resort to the ‘trial and error’ method to find a solution. This can be limited by planning and using human brains a little longer before the experimental phase.
Ruotsalainen also points out that the environmental impact of a data centre can be reduced by establishing the centre in existing buildings. Renovating vacant premises has a lower environmental impact than building from scratch.
Using the brain before AI
The world of researchers and academia is a world of its own, but what can an ordinary tech worker do at their desk to ensure that their use of AI causes as little harm to the environment as possible?
Ruotsalainen's answer is predictable, but it is actually quite comforting from a human perspective.
– Think first. Only then ask the AI.
Simply formulating the problem at hand affects how many clarifying follow-up questions the AI asks and how many detours it ends up taking.
When you take the time to sit down and carefully formulate the question, it often happens that the issue starts to resolve itself simply by writing the question down. You don’t even need AI. So slow thinking can actually be eco-friendly.
“We can each ask ourselves whether every caricature of our own face posted on social media is really necessary,” says Ruotsalainen.
The environmental footprint of artificial intelligence is gradually coming to light
It is likely that, in the near future, calculations of the environmental costs and benefits of artificial intelligence will become considerably more accurate than they are at present.
Researchers at the University of Jyväskylä have been developing calculation methods for the ecological footprint of companies and organisations for several years now, ensuring that the calculations take into account not only carbon emissions but also, for example, land use and other forms of pollution.
A comprehensive ecological footprint has already been calculated for, among others, the S-ryhmä grocery stores, the City of Tampere and the Finnish Academy of Science and Letters, but not yet for artificial intelligence. In principle, it would be possible and interesting, but not easy, suggests Sami El Geneidy, head of the Ecological Footprint Group and researcher at the University of Jyväskylä. El Geneidy has a background in business studies, specialising in corporate environmental management.
The group’s other leader, Professor of Ecology Janne Kotiaho, believes that, globally, AI is currently more commonly used for entertainment than for serious work. The emissions from AI, generated for example by research work, would therefore be smaller for the time being compared to the emissions from entertainment running on mobile phones and laptops.
“Whatever the balance between entertainment and work ends up being, it is possible to reduce emissions from recreational use through mindful behaviour,” Kotiaho points out.