What Is AI Doing to The Planet?
Traditional “Going Green” campaigns tend to evoke the classic activities that are associated with promoting environmental sustainability. Recycle your bottles! Take the bus! Go vegan! These are the heavy hitters when it comes to messaging about saving the planet. But one area of substantial energy use that is less often alluded to in such campaigns concerns the building and sustaining of new technologies like AI. Researchers are increasingly quantifying and framing the costs of using computing power, running required servers, and developing new technologies. Despite this, the environmental costs of using and developing technologies tend to be overlooked in practice. For many, the environmental impact of technologies – from the things we do every day like video streaming to the use of more complex AI or blockchain tools – is just an assumed, but unknown, “cost of doing business”.
In a distributed economy that increasingly hires remote workforces, digital communication and tools are the core of business operations. All the infrastructure that keeps these things running is easy to forget about which, in the pursuit of everyone’s personal convenience and constant connectivity, is kind of the point. Let’s consider the scenario of a project team at a technology firm, working remotely, and let’s say they’re working on a new AI or machine learning model. This is a scenario that unfolds thousands of times across the globe each year.
Just to simply communicate to one another in a remote working environment, a single hour of video conferencing by anyone on the team is enough to generate around a kg of CO2 and require around 12 liters of water. To stream an hour of HD video, estimates of emissions costs range from .036 kg to .055 kg of CO2 per hour. Scaled around the globe, these costs grow very large. These costs are solely from the use of common communication channels and media services; we haven’t yet factored in the costs of computing to train the AI itself.
Researchers in recent years found that training advanced AI models consumes more energy than flying around the world 315 times. Other work on the environmental impact of AI and machine learning expects the costs to grow exponentially with increases in computational power. One of the difficulties with the energy costs of processes like training AI models is that these costs are so widely distributed and indirect. It’s easy to visualize a plastic straw finding its way into a sea-turtles windpipe. By contrast, it’s harder to visualize how technology development negatively contributes to the climate, despite being far more costly.
Countless digital activities contribute negatively to the environment, presenting an ethical trade-off that is starting to be recognized in some, specific fields. Cryptocurrency mining is an area that has been one of the first digital activities to gain attention — and even criticism — for its steep environmental costs.
People’s moral intuitions tend to overlook impacts on the environment because the effects are disassociated from any one individual and occur over long periods of time. Add to this the difficulties in quantifying the energy and emissions impact of AI development, and the environmental impacts of AI are viewed as impersonal and insurmountable. We know that people actually express less moral concern the larger they perceive a problem to be, and environmental impacts certainly fall into that camp. The environmental costs that come with exploring the bounds of AI are significant, yet they will be continually deferred for the sake of innovation.
There are calls to consider AI development, as there are with cryptocurrencies and supercomputing, within a more environmentally sustainable framework. One scholar suggests a government mandated “proportionality framework’” to “…assess whether training or tuning of an AI model for a particular task is proportional to the carbon footprint, and general environmental impact, of that training and/or tuning.” As we’ve seen with ESG initiatives, it’s certainly possible to mandate environmentally conscious development. The question remains whether AI, in its continually increasing use of resources, will fall under this umbrella of scrutiny.
The reality is that environmental hazards and climate change disproportionately affect developing countries. We cannot overlook how the technology economy in richer countries — and all of its extreme environmental impact —is, therefore, developing at the expense of poorer countries. The actions of the world’s leading tech companies, many of which are pioneering powerful AI models and supercomputing applications, have a role to play in achieving a sustainable future. But instead of acknowledging or addressing the real ecological costs of AI development, many of these companies remain unaware of them. AI evaluation and risk management tools tend to be concerned with the direct impacts of models on individuals or ethical values, rather than with indirect and societal impacts like climate change. Calls from the AI ethics community to include sustainability in tools that evaluate AI are an important step in the right direction. But the question remains: will industry listen?