The carbon emissions of writing are lower for AI than for humans
But there's still value if you're writing to prepare for "other kinds of productive human work"
TL;DR - AI writing is “1,100 times less impactful than a US resident writing,” environmentally speaking. If you’re out of ethical arguments against Essay Mills, here’s an environmental one.
The title of the research paper is “The carbon emissions of writing and illustrating are lower for AI than for humans” published in Nature1. The publication is no small deal and while the conclusions of the article (if you’re like me) will give you a little heart burn, they might also help reframe efficiency, AI-use, and how we evaluate the environmental impacts of their use.
We’ve all heard that a single prompt to ChatGPT can use as much as a bottle of water and 0.14 kilowatt-hours of energy2 (that’s for just a 100 word email).
This Nature paper works to normalize data: it takes into account the power and energy used in training, depreciating those values over time across the many queries, including the power usage of your interaction with the chat interface. In order to compare against the human-effort equivalent it uses well established estimates for per capita energy assumption in the US (it also uses data for India, a country with a much lower per capita energy use).
The comparison is normalized over a page of writing (250 words). The average human, the paper estimates, takes about an hour3.
Human output of 250 words? 3,600 seconds & 1400g CO2e
ChatGPT’s time to output 250 words? 3.8 seconds & 2.2g CO2e
Based on the calculations AI writing is “1100 times less impactful than a US resident writing” the same amount.
A few thoughts
Default human CO2 impact: Whether or not we’re writing, we’re burning CO2 because we’re nasty humans requiring clothing, food, heat and housing (it’s 20F outside here in the Mid Atlantic and its toasty inside, so I’m not going to dispute having a constant nonzero environmental impact). That said, being un-alive (e.g. reducing my carbon footprint to zero) is not an option. Plus (as the authors share) “the freed human time may also incur new environmental costs” (e.g. I could go burn trash instead of writing).
Co-authoring: The authors do call out that even AI writing needs editing, so a well written output likely requires a collaboration between AI and human to craft something readable and valuable (increasing the CO2 output to something above the miniscule AI value but well below the un-assisted human value).
Authorship: You’ll be happy to know that the nihilistic take on ecologic impact is tempered by a single paragraph about the value of writing, “human authorship may implicitly be training for other kinds of productive human work that would be lost in the face of the proliferation of AI writing.” Ya think?
In conclusion
As flabbergasted as I was after reading the article, my wife–a physician–was gracious to remind me that not all writing is academic or for learning and that significant portions of professional writing is documentation, note taking, data entry, and transcription. These are tasks not for “training for other kinds of productive human work” but still required to be completed. They are inefficient uses of human time (who wants their doctor spending five hours a night recalling their patient visits when AI could efficiently document and summarize the same thing in a fraction of the time and 1100 times less emissions? No one).
So while I take a lot of issue with the premise of the journal article, I don’t dispute the conclusions or calculations and instead wanted to put these new factoids to work in a way that was productive. Which got me thinking about the inefficiency of contract cheating and essay mills (outsourced human labor). If ever there needed to be an environmental argument against it, this is it.
Cheaters could be courteous enough not to pollute by hiring an essay mill. 🤦
Fun Fact: the bar used here is Mark Twain’s writing speed of about 300 words per hour; he probably could have typed a little faster with a Dvorak setup ;)