Generative AI possesses the ability to create original content and is frequently discussed in relation to artists’ and copyright holders’ rights, as well as labor issues. However, its environmental impact has received far less attention.
As AI becomes an increasingly powerful tool supporting our creative practices, let us reflect on recent research that offers compelling insights into its energy consumption and environmental footprint.
Energy Consumption Varies by Task
Researchers at the AI startup , in collaboration with scientists from Carnegie Mellon University, conducted a study focusing on the carbon dioxide emissions generated by different AI tasks. They precisely measured the emissions produced.
The findings revealed that simple tasks such as text classification emit approximately 0.2 to 0.5 grams of CO₂ per 1,000 queries. In contrast, for the increasingly popular image generation tasks, producing 1,000 images can result in up to 1,000 grams of CO₂ emissions.
In other words, simple tasks like text classification are relatively energy-efficient with low emissions, whereas more complex tasks such as image generation significantly increase both energy consumption and carbon output.
Image by Douglas
From an energy consumption perspective, generating 1,000 images requires 2.907 kWh. While this figure may seem modest, consider that fully charging the battery of a Tesla Model 3 requires 50 kWh—equivalent to generating only about 17,200 images.
With platforms like DALL-E, Midjourney, and Adobe Firefly now producing vast quantities of images, if we assume similar energy demands, the daily energy consumption could rival that needed to charge thousands or even tens of thousands of electric vehicles.
Hidden Environmental Costs Beyond the Obvious
While AI consumes energy during task execution, even greater amounts are required for model training and deployment.
Although it is said to be difficult to precisely calculate the energy costs of model training, it is well known that energy consumption is orders of magnitude higher than for task execution.
For example, training the GPT-3 model—which powers the widely popular ChatGPT and utilizes 175 billion parameters—required 1,287 MWh of energy.
Image by doraseiji
Creativity Generated by AI and Humans
When considering AI from an environmental perspective, issues such as energy consumption and carbon emissions certainly warrant attention.
Yet, AI also brings significant benefits, such as enabling image editing and compositing through simple text prompts, eliminating the need for specialized knowledge or advanced technical skills.
The future development of AI may well depend on how we choose to engage with it.
