Does AI impact the environment?
It's a double edged sword
The impact of AI on the environment is a double edged sword. On the one hand it can be used to improve energy efficiency, on the other, the training of deep learning algorithms consumes ever increasing and energy resources.
Increasing energy use
Schwartz et al, report in 2019 in an academic paper, “Green AI”, that “the computations for deep learning research have been doubling every few months, resulting in an estimated 300,000 times increase from 2012 to 2018”. They categorise environmentaly unfriendly algorithms as “Red AI”, where the emphasis is on increasing accuracy, for example in Natural Language Processing, rather than efficiency (less energy consumption).
A recent article in IEEE Spectrum (Sept. 2021) illustrates the enormous amount of energy needed to train deep learning algorithms. As an example, they describe the challenge of reducing image classification errors using deep learning, estimating that halving the error rate would “need more than 500 times the computational resources”. Given progress over the last decade in reducing error rates, just above 10% on the ImageNet dataset in 2020, it is estimated that the error rate could reduce to 5% by 2025.
To achieve a 5% error rate would cost $100 billion
producing as much carbon emissions as New York city does in a month, according to scholars at the University of Massachusetts Amherst.
This is in contrast to the estimated $35 million that it cost Google subsidiary, DeepMind, to train its system to play Alpha Go.
The Natural Language model created by AI company, Open AI, produced the equivalent of 552 metric tons of CO2 during its training, equivalent to driving 120 passenger cars for a year.
The writers in the Spectrum article conclude that “we must either adapt how we do deep learning or face a future of much slower progress”. So concerned were some researchers about CO2 emissions in Machine Learning (ML), they have produced a CO2 impact calculator for ML practitioners.
Green Energy
Of course large companies like Google can afford to invest in “green energy” production like solar and wind as well as commit to transition to renewable energy sources.
In 2021 renewable energy accounted for only around 28% of the world’s total energy consumption.
Whilst the big companies involved in AI have the scale and financial resources to switch to green energy if they wish, it is much harder for other companies and businesses, seeking to train and use ML for their own applications.
Specialised chips
One way to reduce the energy consumption is to use more specialised AI chips that could cut processing requirements by a factor of 5 and even higher with next generation chips.
Earth resources
AI doesn’t just impact greenhouse gases through energy use, it also affects scarce earth resources and how they are mined, something that most purchasers of AI based Digital Assistants like Alexa, are unaware of. In 2018, Kate Crawford and Vladan Joler wrote an award winning essay, Anatomy of an AI System, accompanied by an informative infographic. Using Amazon’s Echo, Digital Assistant, they describe the impact of such a device, throughout its lifecycle, on human labour, scarce earth resources, energy consumption and human rights.
References
https://spectrum.ieee.org/deep-learning-computational-cost
Jeremy Kahn, A.I.’s carbon footprint is big, but easy to reduce, Google researchers say, Fortune April 22, 2021. [https://fortune.com/2021/04/21/ai-carbon-footprint-reduce-environmental-impact-of-tech-google-research-study/]
https://anatomyof.ai/index.html
https://www.nature.com/articles/s42256-020-0219-9?proof=t
https://mlco2.github.io/impact/
Lacoste et al, Quantifying the Carbon Emissions of Machine Learning, Cornell University, Computers and Society, 2018.[https://arxiv.org/pdf/1910.09700.pdf
Green AI, Schwartz et al, Cornell University, Computers and Society, 2019. [https://arxiv.org/pdf/1907.10597.pdf
Renewable Energy Institute data: https://www.renewable-ei.org/en/statistics/international/