The study was conducted by researchers from the University of California, Irvine, and the MIT Sloan School of Management.
This article is reprinted with the authorization of AI New Media Qubit (public account ID: QbitAI). Please contact the source for reprinting.
AI generation can emit three orders of magnitude less carbon dioxide than human hand-drawing !
The conclusion comes from LeCun’s latest paper shared by “Very Stinging”:
Judging by the number of forwarded comments, netizens can no longer sit still.
Some netizens were a little surprised when they saw this result:
Who would have guessed that generative AI is more creative and environmentally friendly?
There are also a handful of netizens who have not read the paper and asked:
Have you included the carbon dioxide emissions during model training?
In addition, many netizens have questioned the calculation methods used in the paper.
So how is it calculated in this paper? Are training models included? Let's take a look together.
AI vs humans
The study was conducted by researchers from the University of California, Irvine, and the MIT Sloan School of Management.
In this comparison of human and machine carbon dioxide emissions, the AI team members sent by the researchers include ChatGPT (text), BLOOM (text), DALL-E2 (image), and Midjourney (image).
Let’s look at the textual comparison first.
writing comparison
The first problem to be solved first is to define the source of carbon dioxide emissions from AI.
Researchers believe that the two main components are the training emissions of the model and the emissions per query . Training emissions are calculated as a one-time cost amortized to each query.
The model training emission data referenced here is: training GPT-3 emits approximately 552 tons of carbon dioxide equivalent; training BLOOM emits 30 tons of carbon dioxide equivalent.
(CO2 equivalent is a unit of measurement used to compare emissions of different greenhouse gases)
In addition, training the model is not enough. It needs to be continuously trained and optimized later. The researchers here default to fully training the model again every month.
The calculation of emissions for each query is based on the data of ChatGPT emitting approximately 3.82 tons per day and responding to 10,000,000 queries. It is estimated that each query will produce 0.382 grams of carbon dioxide equivalent emissions. In the same way, BLOOM will emit 1.5 grams per query.
Comprehensive training and query emissions, it is finally concluded that ChatGPT emits approximately 2.2 grams of carbon dioxide equivalent per query , and BLOOM emits approximately 1.6 grams of carbon dioxide equivalent per query.
In terms of carbon dioxide emissions from human writing, an article in The Writer magazine once pointed out: Mark Twain could write about 300 words per hour, which can be regarded as the average writing speed of other writers.
Based on the above speed, it is estimated that it takes about 0.8 hours for a person to write 250 words (here counted as 1 page of words).
Then the researchers’ algorithm is as follows:
- The average annual carbon emissions of Americans are about 15 tons of carbon dioxide equivalent, so Americans’ hourly carbon emissions are about 15 tons/8760 hours = 1.7 kilograms of carbon dioxide equivalent. Therefore, the carbon emissions produced by Americans writing 250 words are about 0.8 hours x 1.7 kilograms/hour = 1.4 kilograms of carbon dioxide equivalent, about 1,400 grams.
- Similarly, the average annual carbon emissions of Indians are about 1.9 tons of carbon dioxide equivalent. Then the carbon emissions of Indians are about 0.22 kilograms of carbon dioxide equivalent per hour. Therefore, the carbon emissions produced by an Indian writing 250 words are about 0.18 kilograms of carbon dioxide equivalent, which is about 180 grams.
In addition, based on the average computer power and the carbon emission coefficient of power generation, the researchers also calculated that a laptop that supports human writing produces about 27 grams of carbon dioxide for 0.8 hours, and a desktop computer produces 72 grams.
Let’s look at the comparison results below.
The carbon dioxide equivalent emitted by BLOOM for each page of text generated is 1/1500 of that of American writers and 1/190 of that of Indian writers.
The emissions of ChatGPT are 1/1100 of those written by American writers and 1/130 of those written by Indian writers.
Considering the use of computers, AI writing also produces much less emissions than the total emissions of humans and computers.
Let’s take a look at the image comparison.
Plot comparison
DALL-E2 refers to the method of ChatGPT, and the emission amount per query is about 2.2 grams.
Midjourney CEO David Holz once said that each image requires tens of tera operations.
Therefore, the researchers calculated the power consumption based on the amount of calculations and then converted it into carbon emissions. Finally, they estimated that Midjourney emits approximately 1.9 grams of carbon dioxide equivalent for each image generated.
When calculating the carbon dioxide emissions of human drawing, the researchers first used industry data to calculate that the average drawing project fee for a human illustrator was about $200, and the average hourly wage was about $62.50/hour. Therefore it is estimated that it takes an average of 3.2 hours to complete a commercial drawing project.
The carbon emissions of illustrators in the United States and India are also assessed here.
The average American’s annual carbon emissions are about 15 tons. It takes 3.2 hours to create one image, corresponding to a carbon footprint of approximately 5,500 grams . The average annual carbon emissions of Indians is about 1.9 tons. Drawing a picture corresponds to about 690 grams of carbon emissions .
In addition, computer carbon emissions must also be included in the calculation. Laptops produce about 100 grams of carbon dioxide emissions in 0.8 hours, and desktop computers produce about 280 grams.
Let’s take a look at the comparison results.
The carbon dioxide equivalent emitted by DALL-E2 is about 1/2500 of that of American illustrators and about 1/310 of that of Indian illustrators.
Midjourney emits about 1/2900 of the carbon dioxide equivalent of an American illustrator and about 1/370 of an Indian illustrator.
Also considering computer use, AI drawing has much less emissions than the total emissions of humans and computers.
Other factors should also be considered
Although according to the calculations of the researchers, AI emits much less carbon dioxide during writing and drawing tasks than humans doing manual writing and drawing, they also pointed out its limitations:
Researchers believe that not all fields are suitable for AI intervention, and humans are more efficient in some simple tasks. And future technological changes may alter the extent to which AI and humans impact the environment.
Secondly, AI also has potential social impacts, which may lead to job losses and disputes over the legality of training data.
In addition, as AI technology improves, it is likely to lead to increased demand for goods and services produced by AI, leading to further increases in resource use and pollution through rebound effects.
In short, researchers believe:
AI can play an important role in all areas of society and is not currently trapped in carbon emissions issues. Although the carbon emissions of AI cannot be ignored, AI carbon emissions are currently far lower than humans on certain writing and drawing tasks, and the advantages of AI over humans in carbon emissions should not be ignored.
Caused heated discussion among netizens
The paper sent by LeCun naturally attracted a lot of attention.
Netizens who have read the paper have questioned the calculation method:
There is a big problem with the methodology of this article. It cannot simply compare the emissions of a person and the emissions of an AI model.
Some netizens brought out the Jevons Paradox, believing that as the use of AI increases, carbon dioxide emissions will bounce back:
I've generated far more images with Midjourney than I'd ever want to buy from a human artist; when the price elasticity is less than 1, the reduction in unit cost results in an increase in total expenditures, although I admit I probably won't generate 10,000 images .
In addition to those who complain about the flaws in their methods, some netizens believe that this kind of emissions is insignificant compared to the emissions from other activities:
Who are the biggest emitters of carbon dioxide? Shouldn’t technicians pay more attention to activities that really have a significant impact?
Although the researchers have more or less mentioned these issues in the final discussion of the paper, they have still become a hot topic among netizens.
Of course, there are also netizens who affirm this comparison:
Despite the flaws of this approach, I think it opens the door to discussion and potential new ways to better compare AI and "humans" (traditional counterparts), a topic that must be opened.
Some netizens also sent this wave of Q to Lao Huang (doge):
Don't tell Jensen (Huang Renxun).
