Artificial intelligence training is powered mostly by fossil fuels, according to one of the largest studies of its kind. Less than 25 per cent of AIs use low-carbon energy sources such as hydroelectricity and nuclear power during training.
鈥淧eople see AI to be this intangible thing that lives in the cloud,鈥� says at Hugging Face, a company that develops tools for sharing AI code and data sets. 鈥淏ut behind AI models, there are layers and layers of hardware and tech, all of which have an impact on our planet, either via the energy they consume or the greenhouse gases they emit.鈥�
Luccioni and at the Mila 鈥� Quebec Artificial Intelligence Institute in Canada studied the carbon emissions from 95 high-performing AIs trained on a tasks such as natural language processing and computer vision. By evaluating the location of the computers or data centre cloud servers involved, the researchers were able to determine the main energy source powering each AI.
Advertisement
Of the 95 AIs, 73 drew from electrical grids primarily powered by coal, natural gas or oil. Just 19 relied mainly on hydroelectricity, and 3 were supported by nuclear power. Solar and wind power were not included because they were not the main energy source for any AI training location.
The study also showed that emissions associated with AI training increased 100-fold between 2012 and 2021. That could be exacerbated by the recent trend of tech companies training increasingly huge AI models on growing amounts of data. 鈥淚f the trend is to keep increasing the size, even with a larger contribution of renewable energy, the carbon emissions of such large models will be significant,鈥� says Hernandez-Garcia.
It鈥檚 important to look beyond the yearly average for emissions generated by AI training, because a power grid鈥檚 mix of high-carbon and low-carbon sources can vary from hour to hour, says at the University of Oxford. That means AI training鈥檚 carbon footprint could be wildly different depending on the time of day 鈥� a factor that was not analysed in this latest study.
There are also multiple ways to reduce AI training鈥檚 carbon footprint, says at the University of Florence in Italy. Verdecchia鈥檚 work has shown how modifying the can reduce energy consumption by up to 75 per cent in some cases.
Reference:
