AI and the environment: what impact, what obligations?

AI & Environment: 5 points to remember

  • AI can drive progress on today’s energy and environmental challenges
  • Nevertheless, numerous studies illustrate the extent of AI’s negative impact on the environment, particularly in terms of the energy and resources it consumes.
  • The sustainable impact of an AI system has three pillars: environmental balance, socio-economic contribution and health impact.
  • In order to minimize the environmental cost of AI systems, the use of short supply chains and the reduction of energy consumption can be recommended solutions.
  • In its latest version of June 14, 2023, the AI ACT has incorporated new environmental compliance obligations for AI systems

Artificial intelligence has an ambiguous relationship with the environment. Presented as offering solutions to energy challenges, AI is increasingly the subject of reports warning of its impact on the environment. AI relies on infrastructures that store large amounts of data, and therefore consume a lot of resources and energy. So how can we reconcile AI and environmental protection? We will discuss these new challenges and present the solutions and possible obligations that AI systems will have to comply with under the new European regulation on Artificial Intelligence.

  • The notion of sustainable impact and measuring the energy consumption of an AI system

The notion of sustainable impact must be taken into consideration when exploiting artificial intelligence. Hub France’s AI Ethics white paper identifies three sub-principles for including sustainable impact in the use of AI.

Firstly, the environmental balance, which includes the environmental footprint of an AI system, calculated from the sum of direct, indirect and induced environmental impacts due to its production, development and use. The use of an AI system entails several types of energy consumption, such as the consumption of a GPU – “graphics processing unit” – or the consumption of the servers and the equipment used to run the servers. There is a trend towards initiatives to calculate the energy consumption and carbon footprint of an AI system. Green Algorithms and ML CO2 Impact are just two examples. Criteria such as resource consumption for the process, or power consumption for the datacenter as a whole, are used to perform this calculation.

Secondly, the socio-economic contribution to the sustainable impact of an AI system. The criteria used to assess this contribution are the contribution to collective well-being, the objective of complementarity with human beings and the conditions of access, which must be egalitarian.

Finally, the last principle relates to the impact on health that an AI system can have, for example, in research or the implementation of a public health policy. In this important case, the AI system will have to respect ethical values as well as the protection of personal data.

This article will focus on the first principle concerning environmental impacts. The last two principles will be discussed in other NAAIA articles.

  • The impact of AI on the environment

AI has a significant ecological impact. Cédric Villani, former French Member of Parliament, specialist in mathematical analysis and author of the report on the implementation of a French and European strategy in artificial intelligence, explains this impact as a result of the use of data storage infrastructures necessary for the operation of AI, as well as ” pollution from mining, consumption of resources, energy and even space “.

Various researchers have taken an interest in the subject, and have published studies that help us to understand the extent of the negative effects on the environment. A team from the University of Massachusetts at Amherst, including Emma Strubell, claimed in a 2019 study that certain neural networks used in natural language applications, trained for 4 to 7 days, consumed as much energy as a human in 57 years.

  • AI at the service of energy challenges?

AI can sometimes serve as a solution to today’s climate and energy challenges. According to UNESCO’s working paper on AI for sustainable development, AI can foster advances in ecology and biodiversity research, as well as in ecosystem management. For example, AI can be useful for disaster prevention, as has been tested in a Japanese city, notably for the detection and implementation of a tsunami warning prototype. Freshwater management has also been efficiently monitored using AI based on statistical modelling.

The ecological transition can also use AI to develop innovative, environmentally-friendly solutions. The construction of efficient renewable energy production sites, for example, can be identified by AI to find the most favorable sites in terms of exposure to sun, climate or wind.

  • What solutions for responsible AI?

To limit the pollution generated by artificial intelligence, researchers are recommending that sustainable impact criteria be taken into account. The aim would be to minimize the environmental cost of the production chain for the materials needed to create systems using AI, such as the use of short circuits. When it comes to developing and operating AI systems, one solution could be to reduce the energy required to minimize the associated environmental costs.

Finally, in order to offset the environmental cost of the production chain, development and operation, it would be interesting to promote uses with a positive impact on the environment.

According to Cédric Villani, ” Responsible AI requires responsible manufacturing, implementation and control processes. The responsibility lies with mankind. “. It recommends questioning ” the purpose and architecture of algorithmic systems: their transparency, their a posteriori evaluation, their biases and their responses to environmental issues”. Digital sobriety is also put forward, and would consist in limiting the uses of AI through good practices to be adopted.

  • European AI regulations: the “social and environmental well-being” principle

The European Parliament’s amendments of June 14 – summarized in a Naaia article – emphasize one of the main principles of the AI ACT, namely social and environmental well-being. New environmental provisions have been incorporated. In the newly-created recital 46a of the AI ACT, the European Parliament points out that ” AI systems can have a significant environmental impact and high energy consumption during their life cycle “.

The new recital 28a, for its part, supports the idea of the “fundamental right to a level of environmental protection”.which must be “taken into account when assessing the seriousness of the harm that an AI system may cause, particularly with regard to the consequences (…) for the environment”.. According to these amendments, an AI system may be considered “high risk” if it “presents a significant risk of harm to the environment”.

More than high-risk systems, all AI systems will, according to the new Article 4 bis of the AI ACT, be “developed and used in a sustainable and environmentally friendly manner”.

The AI ACT thus intends to impose obligations to respect the environment when setting up an AI system.

Although artificial intelligence can contribute to the energy transition, it also presents a number of environmental challenges. Companies therefore need to integrate the idea of sustainable impact into their AI systems, as well as the idea of digital sobriety. The European Union’s AI ACT aims to include environmental obligations as key principles to be met by the various artificial intelligence systems operating in Europe.