Researchers discover a method to reduce AI carbon dioxide emissions

Bold Claims and Serious Flaws: AI's Environmental Impact

A recent study from MIT, the University of California, and other institutions has sparked controversy with its headline: "AI Writing and Drawing Have Lower Carbon Emissions Compared to Humans." While this finding may seem impressive at first glance, a closer examination reveals some serious flaws that undermine the credibility of the study.

The study compares the carbon emissions of AI systems (including ChatGPT, BLOOM, DALL-E2, and Midjourney) with those of human writers and artists. It concludes that AI emits significantly less carbon dioxide equivalent (CO2e) when performing tasks such as writing and illustrating. For example, ChatGPT emits only 2.2 grams of CO2e per query, while human writers from the United States are responsible for approximately 1400 grams of CO2e per page.

Yann LeCun, Chief Scientist at Meta AI, shared this study on X, and many people have been discussing the issues it raises and the many things it overlooks.

The study overlooks the high energy consumption involved in training AI models. AI models like ChatGPT and DALL-E2 require significant computational power and electricity during the training phase. The carbon emissions associated with training these models are substantial but are often ignored in such analyses. Considering the entire lifecycle from training to operational use is necessary when evaluating the environmental impact of AI systems.

Interesting Perspectives

One issue with the study is the method used. It assumes that an individual's emissions while writing are consistent with their overall annual carbon footprint, equating one hour of writing to one hour of breathing. This assumption is far from scientific and fails to consider that individuals continue to breathe even when they are not actively writing.

A user on HackerNews explains that this is akin to attributing the emissions of a person riding a bicycle solely to the act of cycling, disregarding the carbon emissions generated by their daily existence. In short, humans continue to breathe whether or not they are writing. On the other hand, AI systems can be turned off when not in use, effectively reducing their carbon footprint to the tasks they perform. This distorted comparison fails to accurately consider the comprehensive emissions of humans and AI.

Another user points out that the study also overlooks a crucial aspect of AI development: data. AI systems rely on large amounts of text and image datasets to generate content. These datasets, created by humans, require energy and resources for data collection, labeling, and preparation, which should be included in environmental considerations.

The study also neglects the ongoing improvement and fine-tuning of AI models. These models require regular updates and adjustments, consuming computational resources and generating carbon emissions. By focusing on a page of text or an illustration, the study fails to consider the operational costs of AI systems, resulting in an incomplete picture.

In All Honesty, It Doesn't Matter

In essence, AI models are built on the foundation of human creative work, so directly comparing AI emissions to human emissions without acknowledging this significant contribution is misleading. In response to these valid criticisms of the paper, some may argue that even if AI systems emit more CO2 throughout their lifecycle, it doesn't matter because they are more efficient in certain tasks.

In conclusion, while the initial results of this study may suggest that AI is a more environmentally friendly alternative to human creativity, further research reveals flaws in its methodology and an incomplete picture. However, AI researchers are not ignoring the environmental impact of AI and are actively taking measures, such as investing in nuclear energy, to address this issue.