Renowned AI ethicist Mark Coeckelbergh proposes ‘AI for Climate’ suggesting that we use AI “for dealing with environmental and climate problems” [8, 5]. Given my earlier distinction of sustainable AI, I would label this as AI for sustainability. And while I couldn’t agree more, I also believe it is necessary to focus our attention on the sustainability of AI. This change in framing is paramount as it means that one can no longer talk about AI for Climate or AI4Good without at the same time addressing the impact that developing a particular AI model will have on environmental sustainability.
Thankfully there are a select few studying and developing AI models who are already bringing attention to the issue and outlining areas in need of further study. In a 2019 paper by Strubell et al. , it is argued that there are both financial and environmental costs to Deep Learning (DL) models for natural language processing (NLP). The financial costs were attributed to hardware and electricity or cloud compute time (which raised ethical issues as to who has access to such hardware etc.) whereas the environmental costs were attributed to “the carbon footprint required to fuel modern tensor processing hardware” [17, 1]. The authors acknowledge the energy required to power the hardware for training such models is considerable given that training happens over the course of weeks or even months .
The authors also point out that while it is possible to obtain some of the required energy from “renewable or carbon credit-offset resources, the high energy demands of these models are still a concern since (1) energy is not currently derived from carbon–neutral sources in many locations, and (2) when renewable energy is available, it is still limited to the equipment we have to produce and store it, and energy spent training a neural network might better be allocated to heating a family’s home” [17, 1]. This last point should strike a chord for all of us; as said before, there are approx 600 million people in the world without access to modern electicity and instead of prioritizing the provision of electricity to these homes, we are prioritizing the training of AI models that can beat the world champion at the game of Go (AlphaGo). It is time for such calculations to be made explicit and to be evaluated in an open forum.
The Strubbel et al.  paper goes on to show that ‘tuning’ (aka re-purposing or refining) an AI model is more expensive than training a model to begin with. This kind of finding is crucial for the policy makers to understand to make decisions concerning the proportionality of certain AI methods compared with its intended application. Meaning, it is time for policy makers to govern AI at a more detailed level and suggest that certain methods, for example tuning a deep learning NLP model, should not be permitted for ethically charged tasks like recruitment of new employees or prediction of employees who may be on the verge of quitting. The reason being that the costs to environmental sustainability are simply too great to justify such a menial (not to mention ethically problematic) application. This could also cause society to pause when a particular AI model will be used in an application context which will require constant tuning. As society evolves in its communication, transportation, and social habits, old AI models will need constant tuning to continue to be effective. These costs must be added to any proportionality calculation.
One of the final recommendations from Strubbel et al.Footnote 3 is that “authors should report training time and sensitivity to hyperparameters” as “this will enable direct comparison across models” (Strubell, Ganesh, and McCallum 2019, 5). To date, there are two possible tools available for calculating emissions: the ‘Machine Learning Emissions Calculator’ for estimating the carbon footprint of GPU compute through specifying hardware type, hours used, cloud provider, and region , Anthony et al. ; and, the ‘experiment-impact-tracker’ framework “for tracking real-time energy consumption and carbon emissions, as well as generating standardized online appendices” [10, 1]. Each of these approaches aims at the ‘mitigation of carbon emissions and reduction of energy consumption’ to facilitate the sustainable development of ML .
If we recall from the introduction, studies have shown that ‘Google’s AlphaGo Zero generated 96 tonnes of CO2 over 40 days of research training which amounts to 1000 h of air travel or a carbon footprint of 23 American homes’ [1, 15]. Or, training one large NLP model (aka a transformer), with neural architecture search, resulted in over 600,000 CO2e(lbs), roughly the equivalent of carbon emissions of five cars (over the lifetime of the car) . These numbers are overwhelming to read. What’s worse is that we have only a few studies to call on to learn numbers like this. In other words, we need more studies to fully grasp the extent to which these findings can be supported or refuted. With tools like the ‘machine learning emissions calculator’  and the ‘experiment-impact-tracker’  this should no longer be the case. This means the tools to track carbon emissions are there, however, greater incentives are needed to encourage researchers and industry developers to measure and report such findings.
Building on the work of Henderson et al., Anthony et al. propose ‘carbontracker’, “a tool for tracking and predicting the energy consumption and carbon emissions of training DL models” [3, 2]. Not only does the ‘carbontracker’ tool allow for the generation of carbon impact statements but it also allows for the model training to be “stopped, at the user’s discretion, if the predicted environmental cost is exceeded” [3, 2]. Thus, the ‘carbontracker’ provides the possibility that if a training exceeds a responsible use of energy consumption, or generation of carbon emissions, training of the model can be stopped. Again, this is the kind of tool that should be known to policy makers to create governance mechanisms for limiting the amount of carbon emissions, with the tools to end the training when emissions reach an unacceptable threshold.
In short, while the use of AI for achieving sustainability is to be applauded there are many reasons for which the environmental costs of AI, the sustainability of AI, need to be studied and made transparent to the AI community, consumers, and policy makers. More to the point, “the carbon emissions that occur when training DL models are not irreducible and do not have to simply be the cost of progress within DL” [3, 3].