The Carbon Emissions of Writing and Illustrating Are Lower for AI than for Humans – review
This is a discussion of the results of the paper “The carbon emissions of writing and illustrating are lower for AI than for humans”
An interesting study recently made the rounds on the internet. Its title is “The Carbon Emissions of Writing and Illustrating Are Lower for AI than for Humans” [1] and it argues that for some tasks, generative AI is more energy efficient than humans.
In essence, the authors estimate the carbon footprint of ChatGPT writing 250 words (one page) versus a human writing them.
In this blog post, I comment on their methodology and findings.
My review will focus on the part of the paper that compares human emission to AI emissions wrt writing tasks and will ignore the section about illustrating.
Review of estimates of ChatGPT’s electricity use and carbon footprint
Let’s first take a look at the estimates of ChatGPT’s energy use and carbon footprint presented in the paper.
The electricity consumption their comparison relies on are what they call “informal estimates” that stem from an online blog post. With all due respect to the author of that blog post, the estimate does not account for several factors and could therefore be too low. Some factors not accounted for:
- ChatGPT’s ongoing networking electricity consumption
- ChatGPT’s ongoing CPU consumption
- Embodied emissions of the hardware that ChatGPT is running on
- Embodied emissions of the data center building
- The emissions from all the thousands of hours of human work that has gone into developing ChatGPT
I think it’s a shame that a research paper doesn’t rely on a more comprehensive analysis of ChatGPT’s energy consumption. If one such analysis did not exist at the time of writing, the authors should have made the effort to make one.
In addition to the factors above, the authors don’t account for human time spent writing a prompt, and they assume that it only takes one prompt to get the text right which might on average be a bit unrealistic.
Review of methodology used to calculate human emissions
Let’s now analyze how the authors estimate the carbon footprint from a human writing 250 words. This is in my view the most problematic part of the paper.
What the authors do is this: 250 words takes approximately 0.8 hours to write, and the average American emits 15 tons CO2e in a year, so an American writing 250 words has a carbon footprint of 15 tons / (365*24) * 0.8. That’s roughly 1.4 kg CO2e.
I take issue with this approach for several reasons. The 15 tons is obtained by dividing the total carbon footprint of the US and dividing by the number of citizens in the US. The total carbon footprint of the US is defined as the emissions from burning fossil fuels and from certain industrial processes such as cement and steel production. As such, the 15 tons figure does not actually reflect the emissions of individual behavior and can therefore not be used to compute a human’s carbon footprint from performing any task.
In addition, when using this method, a human’s carbon footprint would be the same if the human just sat idle. Even if using generative AI, that human’s carbon footprint would be the same. In fact, the carbon footprint of a human writing a page might in theory increase from others’ use of generative AI if that generative AI is powered by fossil fuels.
So, with this method, the use of ChatGPT does not reduce anyone’s carbon footprint. A better method would be one method with which you could measure a reduction in human carbon footprint from using generative AI.
In addition, the authors include laptop energy consumption when calculating human carbon footprint, but not when calculating ChatGPT’s carbon footprint.
Conclusion
To sum up, the paper poses an interesting question, namely to what extent it is more energy efficient to have AI carry out a task. But almost no matter how polluting the AI, the answer will tend to favor AI when a human’s carbon footprint from carrying out a task is calculated the way this paper does.
What do you think of the paper?
That’s it! I hope you enjoyed the story. Let me know what you think!
Get the benefits of Medium and support my writing by signing up for a Medium membership HERE.
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://arxiv.org/abs/2303.06219
