Is Generative AI Worth Its Environmental Footprint?
Generative AI might have a noticeable environmental footprint, and this story discusses what we may stand to gain in return.
Generative AI is receiving a lot of attention these days. ChatGPT reportedly has hundreds of millions of users and similar functionality is allegedly being built into a multitude of digital products ranging from Microsoft Word and Teams to search engines.
The environmental footprint of generative AI could be noticeable if billions of people begin to use generative AI extensively on a daily basis [1].
But will the value we accrue from this technology outweigh the potential environmental costs?
That’s the question I will try to shed some light on in this article by outlining some perspectives on what we may stand to gain.
First, I’ll provide some perspectives on potential productivity gains from generative AI.
Then, I’ll discuss whether generative AI will be a net positive or a net negative technology.
Next, I’ll discuss the extent to which generative AI may reduce inequality.
Finally, I’ll provide my perspective on whether we should expect generative AI to speed up the green transition.
Scope
Please note that what follows concerns only generative AI. Generative AI is distinct from other types of AI.
Also note that I focus solely on generative AI that generates text (including code) and discard image and sound generation technology.
For a simple working definition of generative AI, see [2].
Will generative AI make us more productive?
Let’s consider the extent to which generative AI could lead to productivity gains. One study of the effect of ChatGPT on productivity in various writing and analysis tasks found that those who used ChatGPT spent 40% less time on the tasks than the control group that didn’t use ChatGPT. In addition, as per peer reviews, the quality of the solutions to the tasks were 18% higher [3].
A survey among software developers showed that 88% of the survey respondents reported that they experience that using the generative AI tool GitHub Co-Pilot increases their productivity [4]. The results were echoed in a research paper where productivity was actually measured empirically. This paper showed a producticity gain of 126% when using Co-Pilot to implement a Javascript server [5].
As pointed out by Nielsen [6] , these numbers are staggering compared to the average annual labor productivity increase (defined as the value created by a worker per hour), which for the US was 1.4% and for the EU was 0.8% between 2007 and 2019.
But will these results apply outside of the experimental setting of a research study? And can we expect similar productivity gains across the board? For instance, is the task of implementing a Javascript server (the task that Co-pilot helped finish 124% faster) a representative task in software development? I don’t know as I’ve never done that task, but I’m quite sure the productivity gain will not be the same for all programming tasks. I expect productivity gains like the aforementioned can be gained when solving clear cut programming tasks that include writing a lot of boiler plate code, but for now I expect generative AI will only aid minimally – if at all – for tasks where you have to spend a lot of time contemplating the problem before you even know what questions to ask of generative AI. As such I expect the average productivity gain for programming tasks to be lower than 126%.
As generative AI tools – and our ability to use them – become more sophisticated, we might see higher and more widespread productivity gains.
That leaves the question: what environmental cost is society willing to pay for such productivity gains? One way to determine what would be an acceptable cost would be to look at the carbon costs of historical productivity gains. I’ll leave that as an exercise for the reader.
Will generative AI be a net positive technology?
Let’s now discuss different perspectives on whether generative AI will be a net positive technology – i.e. a technology whose footprint is outweighed by emissions reductions that it leads to.
One argument against widespread adoption of generative AI is this: If generative AI does not lead to any reduction in greenhouse gas emissions, it will be a net negative technology – a technology whose use emits more than it reduces. And due to climate change caused by greenhouse gas emissions, we’re supposed to reduce our global carbon footprint – not increase it.
But will generative AI be a net negative or a net positive? It’s easy to imagine that it can save consumption in some places. For instance, it might reduce our need for searching Google and loading websites. But the question is what the carbon footprint of making one Google search and loading a web page is. That’s a complex question to answer.
One number does float around the internet wrt the electricity consumption of making a single Google search, but as far as I know that number is from 2009. Much has probably changed since, and I imagine Google searches today are powered by more sophisticated and more energy consuming machine learning methods. I do, however, expect Google searches to use less energy than a ChatGPT query, although I would venture the difference is less than the factor 10 stated some places [7].
In light of this, I suspect not all generative AI queries will save enough Googling to make up for their environmental cost.
I’ve written a lot about the environmental impact of AI, and one of the comments I frequently get wrt generative AI is that it will reduce our overall resource consumption and carbon footprint. Aside from the case above, I continue to have a hard time seeing how that would be the case. But please challenge this.
A paper titled “The Carbon Emissions of Writing and Illustrating Are Lower for AI than for Humans” [8] was recently published. It argues that generative AI is more energy efficient than humans. I will analyze the claims in this paper in an upcoming blog post.
Will generative AI reduce inequality?
In my discussions with people about the environmental impact of generative AI, I’ve heard some argue that the environmental impact will be worthwhile because they expect the technology will help reduce inequality. Whether we can expect that to be the case is not clear cut. Here’s why.
The aforementioned study of ChatGPT’s effects on productivity [3] found that less skilled people had a larger productivity gain from using ChatGPT than more skilled people. We might therefore expect that access to generative AI can help lift people who are affected by structural inequality.
But there’s disparity between groups of workers in terms of how much generative AI will boost their productivity. Let’s use journalists and carpenters as examples. For journalists who write, the production of their primary output (articles) can be directly supported by generative AI. Generative AI won’t directly aid carpenters in building houses. So it seems that not all groups of workers will benefit equally from generative AI. If the productivity gains from using generative AI lead to higher wages, we might see increased inequality between groups of workers.
This notion is supported by research from MIT where an economist has found that 50 to 70% of the growth in US wage inequality between 1980 and 2016 was caused by automation [9].
At the same time, the International Monetary Fund (IMF) points out that AI could widen the gap between rich and poor nations [10].
In sum, the above studies suggest that generative AI might help decrease skill level inequality wrt tasks that generative AI can assist in, but at the same time, it might increase inequality between groups of workers and between nations.
Will generative AI speed up the green transition?
Now let’s consider the extent to which generative AI could speed up the green transition. It’s obvious that as white collar workers, researchers in green energy and related fields might become more productive as a result of using generative AI. Like others, they might write research articles or code faster. But for the foreseeable future, I personally don’t expect generative AI systems like ChatGPT to contribute directly to the green transition. I have a hard time picturing ChatGPT – a generalised text generation tool trained to come up with the next most probable word based on existing texts – coming up with a breakthrough idea. But I hope I’m wrong. Let me know if you have a different perspective.
I do, however, believe that other types of AI can help speed up the green transition or fight climate change. For instance, other types of AI can help reduce energy consumption in buildings, fight deforestation, optimize shipping routes to reduce energy consumption etc.
Conclusion
Generative AI is on the rise which has led to debates about the environmental costs of this technology. The purpose of this blog post was to shed light on what we stand to gain from generative AI, so as to inform a debate about whether generative AI is worth its environmental footprint.
Initial studies suggest that the use generative AI can lead to productivity gains for some tasks, but it’s still unclear if these findings will generalize to other settings.
Research also suggests that generative AI can help reduce skill inequality between workers within the same occupation, but I argue that it could contribute to rising inequality between workers of different occupations. Further, AI and automation technologies have previously been found to increase economic disparity between nations.
Wrt whether generative AI will speed up the green transition, my personal perspective is sceptical.
In sum, it’s clear that generative AI can have positive effects on productivity, and maybe even on inequality, but it comes at an environmental cost. The question is: do the benefits of generative AI outweigh the costs? That’s essentially a value judgement – one I hope you’re better equipped to ponder now that you’ve read this blog post.
Thanks for reading.
That’s it! I hope you enjoyed the story. Let me know what you think!
Follow me for more on AI and sustainability and subscribe to get my stories via email when I publish.
I also sometimes write about time series forecasting.
And feel free to connect on LinkedIn.
References
[1] https://towardsdatascience.com/environmental-impact-of-ubiquitous-generative-ai-9e061bac6800
[2] https://www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-generative-ai
[3] https://news.mit.edu/2023/study-finds-chatgpt-boosts-worker-productivity-writing-0714
[4] https://github.blog/2022-09-07-research-quantifying-github-copilots-impact-on-developer-productivity-and-happiness/
[5] https://arxiv.org/abs/2302.06590
[6] https://www.nngroup.com/articles/ai-tools-productivity-gains/
[7] https://www.cell.com/joule/fulltext/S2542-4351(23)00365-3?_returnURL=https%3A%2F%2Flinkinghub.elsevier.com%2Fretrieve%2Fpii%2FS2542435123003653%3Fshowall%3Dtrue
[8] https://arxiv.org/abs/2303.06219
[9] https://www.technologyreview.com/2022/04/19/1049378/ai-inequality-problem/
[10] https://www.imf.org/en/Blogs/Articles/2020/12/02/blog-how-artificial-intelligence-could-widen-the-gap-between-rich-and-poor-nations
