A study from earlier this year asserts that artificial intelligence-based systems like ChatGPT, BLOOM, DALL-E2, and Midjourney can create literary and artistic works with lower carbon emissions than humans. Moreover, the claimed difference is far from negligible, with the paper suggesting that AI can produce useful texts and images while emitting 3 to 4 orders of magnitude less CO2 than a human would working manually or with the help of a computer.
AI has become far more capable in recent months, moving from being a buzzword to a truly meaningful and impactful technology. Some of those impacts have been negative, especially as the nascent tech has teething troubles with potential copyright infringement, amplifying some of the worst of human evils, or chewing through scarce resources.
The study, entitled “The Carbon Emissions of Writing and Illustrating Are Lower for AI than for Humans,” looks closely at two popular tasks for AI: AI writing and AI illustration. It comes from authors at the University of California, Irvine, MIT, the University of Kansas School of Law, and more.
AI versus Humans: Writing Tasks
The authors have published the data to come up with its comparisons. For the AI, it took into account the energy consumed during the training phase, as well as operational energy use. To weigh human carbon footprints for the same tasks, it considered the average individual’s emissions for the US and India, and tallied the energy from any computer time separately.
Another assumption made was “the quality of writing produced by AI is sufficient for whatever task may be at hand.” The key findings here may be that BLOOM was 1,500 times less impactful than a US resident creating one page of text, meanwhile the popular ChatGPT could complete the work in a 1,100 less impactful manner.
AI versus Humans: Illustration Tasks
The researchers noted here that, as it is based upon GPT-3, “DALL-E2’s footprint is similar to the footprint of ChatGPT calculated above,” at about 2.2 grams of CO2 per query. Meanwhile, Midjourney was estimated to be responsible for 1.9 grams of CO2 emissions per query.
The researchers propose that “AI image creation produces 310-2,900 times less CO2 emissions per image than human creators.”
The discussion of the paper attempts to address questions that the data may bring up. Firstly, there are a lot of big assumptions behind the comparison numbers. The researchers are well aware of this: atypical results can be had when tasks range from drawing a stick-man at the simple end of the spectrum, or writing “an in-depth, heavily-referenced, original article on a niche scientific topic [which] is currently beyond the capabilities of an AI.” In both of those cases, humans can’t be beaten by current AIs.
There is also the underlying issue of humans having carbon footprints whether or not they create literary or artistic works – or not. Moreover, if a human isn’t writing, for example, they might have a greater carbon footprint from an alternative activity they undertake. Like, if this writer wasn’t ‘crafting’ this article, they might be playing a guitar through a fully cranked stack of 100 watt amplifiers, overclocking their RTX 4090 to within an inch of its life (I can dream), or heating up a kiln to several hundred degrees to fire some home-made pottery.
Some might also find the comparison between a human completing a task against an AI to be distasteful. We’ll see if these comparisons hold up as AI systems get more complex and use more sophisticated technologies.