Artificial Intelligence is seen as a key technology of the 21st century. It can serve countless practical purposes: translations, medical diagnoses, personalized product recommendations and much more. As such, it is likely that AI will gradually enter nearly all areas of society, not only in the form of new products and services, but also through the improvement of existing processes, making them “smarter.” The increasing saturation of society with AI-based solutions, however, means that global energy consumption will increase, not only because of end device usage, like smartphones. A significant portion of the energy consumption associated with AI applications takes place externally, through data transfers and at data centers. It is true that data-center energy efficiency is increasing steadily. But if the sector remains largely unregulated, a rebound effect could set in, meaning the cost reductions associated with energy savings could lead to more intense usage, and therefore to a growth in absolute energy consumption. The European Commission estimates that data-center energy consumption in the EU will increase from 77 terawatt hours (2.7 percent of overall electricity consumption) in 2018 to 99 terawatt hours (3.2 percent of total energy consumption) in 2030. In short: The growing number of data transfers and rising amount of data processing translate into more energy consumption.
It is thus important to consider how AI solutions can be developed, controlled and used as energy efficiently as possible. One potential approach is to educate consumers about AI’s ecological footprint. Certificates and labels like those the European Union provides for “Green IT” could help consumers recognize and choose environmentally friendly options. It’s also feasible to require so-called “carbon impact assessments” for the development and sale of AI solutions and hardware. If the ecological footprint of AI applications is quantified and presented publicly in this way, it could influence consumer behavior.
But it is unclear how effective information is on its own. Not only does this create yet another area where consumers must learn to make informed choices, but experiences in other sectors have also shown that the combination of sufficient knowledge and an environmentally friendly disposition does not necessarily lead to a change in behavior. And energy-efficient applications are unlikely to catch on if they are more expensive or have lower performance, as shown by the results of the study we conducted called “Consumers are willing to pay a price for explainable, but not for green AI. Evidence from a choice- based conjoint analysis.” Especially when it comes to free systems, people are unlikely to accept fees for greater energy efficiency stand- ards. Environmentally friendly AI is therefore unlikely to establish itself on the market based on information and labels alone. That means that political action and government regulation will need to be implemented at points prior to consumer decision-making, as the following graphic shows.
Financial instruments can focus on the marketing stage and regulate supply and demand through pricing. A relatively simple and easily implementable solution would be the inclusion of both producers of AI applications and products as well as data centers in the existing CO2 emissions trading system. This would result in stronger incentives for the development and usage of energy efficient applications and infrastructures for information. By making externally consumed energy and the resulting emissions more expensive, companies would be driven to consider environmental sustainability at the development stage (“Green AI”) instead of only looking at performance (“Red AI”). Such measures can make an important difference. An OECD report shows that different technical decisions – regarding the choice of model, hardware and data center, including their locations – can lead to enormous energy savings.
Finally, regulation can also focus on the development stage of AI applications. This could happen, for example, through the top-runner approach as used in Japan: All providers need to reach the highest energy efficiency standards of the leading provider within a predetermined timespan. Otherwise, they risk being prevented from offering their products on the market – through a government ban, for example. This kind of hard regulation sets new standards by dynamically linking market competition to the respective state of the technology.
Ideally, the use of all of these instruments would complement each other and would be accompanied by the conttinuous collection and evaluation of data. In that effort state regulators should not be forced to rely exclusively on reports and information supplied by the companies themselves. The government must also develop the capacity to effectively monitor companies and, should it become necessary, to intervene.