The carbon footprint of AI:
What are the true costs?
The use of artificial intelligence (AI) is increasing each year across an increasing number of domains (business activities, artistic creation tools, weather prediction, etc.) in our society. The monetary costs of deployment vary with the economic models used by each organization.
However, there exists a hidden cost of which few people are aware. This is the environmental cost (carbon dioxide [CO2] emissions) of training an AI model.
For example, training the GPT-3 model, depending on the continent, can result in carbon emissions levels that vary from single to double (Taddeo M. and al., Artificial Intelligence and the Climate Emergency: Opportunities, Challenges, and Recommendations, June 8, 2021).
Figure 1—Environmental costs (in kg of CO2eq) of a single training run of GPT-3 across different compute regions.
Source: Artificial Intelligence and the Climate Emergency: Opportunities, Challenges, and Recommendations
The carbon footprint of different AI models is not necessarily taken into account during their design. This was observed by a research team. And this same team has developed a framework and a series of measures to be applied to, if not stop the development of AI models, at least reduce carbon emissions (Henderson P. and al., Towards the Systematic Reporting of the Energy and Carbon Footprints of Machine Learning, The Journal of Machine Learning Research, Volume 21, Issue 1, Article No: 248, pp 10039–10081).
However, there are applications to optimize the performance of AI models (e.g., CodeCarbon). There are also calculators that predict the carbon footprint of an AI project. For example, there is the Green Algorithms project (http://calculator.green-algorithms.org) developed by Loïc Lannelongue, Jason Grealey and Michael Inouyefrom the University of Cambridge (United Kingdom) and the Baker Heart and Diabetes Research Institute (Melbourne, Australia).
To calculate the carbon footprint cost, several factors need to be considered for the calculation of an AI model's carbon footprint. These factors include the algorithm design (is the AI model code efficient?), electricity consumption (a factor that varies depending on locality and its production mode), test duration (on average, tests last 24 hours), and server hosting type (Cloud carbon footprint: Do Amazon, Microsoft, and Google have their head in the clouds?, Carbon4, 2022).
In conclusion, calculating the carbon footprint is primarily an overall estimation rather than real results (Mesurer l’empreinte environnementale du numérique, un vrai casse-tête, Next, Mathilde Saliou, 2023).
Regarding the availability of scientific publications on the CO2 cost (and generally, the environmental cost) of AI models, they are rather scarce, as noted by researchers Roy Schwartz, Jesse Dodge, Noah A. Smith, Oren Etzioni (Schwartz R. and al., Green AI, 2019). The research team found that between 2012 and 2019, there were more scientific publications focusing on the accuracy of AI model calculations than on the efficiency of AI models.
Figure 2—Between 2012 and 20,219, AI papers tend to target accuracy rather than efficiency. The figure shows the proportion of papers that target accuracy, efficiency, both or other from a sample of 60 papers from top AI conferences.
Source: Schwartz R. and al., Green AI, 2019.
In 2024, this disproportion in scientific publications still persists. What's most ironic is the implementation of a carbon footprint monitoring system based on Deep Learning and Machine Learning techniques (Carbon Footprint Monitoring System Using Machine Learning and Deep Learning Techniques, IEEE Xplore, 2023).