Use and acceptance of Artificial Intelligence (AI) has expanded rapidly in recent years. It can be found everywhere: in healthcare, education, the workplace... It is almost impossible to find anyone these days who doesn't admit to dabbling in AI in one form or another. But whilst AI promises great rewards, it may come at a great cost: the environment.
An environmental issue
From beginning to end, AI seems destined to be an
environmental nightmare. The materials needed to make the
high-power chips needed for AI are often mined in environmentally
destructive ways, and more of these materials are required than for
conventional microchips. Datasets and models for AI are becoming
more complex, leading to increased electrical energy to train and
run those AI models. In fact, it is estimated that the amount of
computing power used for deep learning increased 300,000-fold
between 2012 and 2018, a far cry from Moore's law, which
observes that overall processing power for computers generally
doubles every two years. Furthermore, estimates show that ChatGPT
requests require ten times the amount of electrical energy that a
Google search does. Other AI chatbots and AI-powered searches fare
even worse.
On top of this, the microchips used for AI applications require greater cooling compared to traditional microchips. Datacentres already require enormous amounts of water to function effectively, and AI datacentres are set to be worse. According to one estimate, the water needed to meet global AI demand is projected to be more than half the water annual water withdrawal of the United Kingdom.
Finally, the rapid expansion of AI is making not only traditional hardware obsolete more quickly, but also older AI technology. This both furthers additional production of high-power chips (to keep up with the edge of innovation), as well as increasing electronic waste ("e-waste"). E-waste contributes to the release of hazardous substances into the environment when not properly disposed of, not to mention that improper disposal wastes valuable resources embedded in the hardware. A recent 2024 report estimates that of the projected 82 billion kilograms of e-waste to be generated in 2030, formal collection and recycling rates will decline to 20% (down from 22.3% in 2022).
How can we square all this with the general consensus that we should both individually and collectively be aiming to reduce our carbon footprint as global warming becomes more and more of a reality?
An environmental solution
More than anything else, AI excels at identifying patterns
and anomalies in data. This is highly useful for refining
operations in a way and on a timescale that humans just can't
manage. Accordingly, despite its environmental impact, AI is also
being implemented in a variety of sectors to save
energy.
For instance, the EPO's recent report on advances in photovoltaics notes that AI has been used to minimise downtime and extend the lifetime of photovoltaics systems. This is critical for maximising solar energy use. AI has also been implemented in power grid systems to predict supply and demand needs to make more efficient use of the available energy. This is particularly useful for renewable technologies that rely on outside, ever-changing variables, such as wind and solar power. Furthermore, AI-powered robots are being considered to assist with classification and collection of e-waste, with the hope that this could increase recycling rates.
Thus, whilst AI has its environmental issues, it could also hold the key to saving the environment too.
The future of AI and sustainability
Given AI's ability to assist in the fight against
climate change, the aim now is to reduce AI's own power
consumption as far as possible without losing its effectiveness. A
recent report published by UNESCO and UCL suggests that small
changes to the way that AI is trained could be the solution. For
example, using smaller, more specific AI models (rather than large,
all-purpose models) can reduce energy consumption by up to 90%,
whilst also providing more accurate results. Other options include
using shorter user prompts and providing shorter model responses,
and rounding down numbers used in the AI's internal
calculations.
Thus, as interest in AI continues to grow, it seems that it will become more and more important for companies to refine their models and focus on using smaller models better suited to specific tasks. Otherwise, the increasing environmental costs of AI, particularly where existing systems can address user needs just as well (if not better), may just fail to justify its use.
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
