[Paper Review] The Carbon Emissions of Writing and Illustrating Are Lower for AI than for Humans

Bill Tomlinson, Rebecca W. Black+2 more|arXiv (Cornell University)|Mar 8, 2023

Energy, Environment, and Transportation Policies4 citations

TL;DR

This study compares carbon emissions from AI systems (ChatGPT, BLOOM, DALL-E2, Midjourney) and humans performing writing and illustration tasks. It finds that AI emits 130 to 1500 times less CO2e per page of text and 310 to 2900 times less per image than humans, indicating significant environmental advantages for AI in content creation despite limitations in social and ethical considerations.

ABSTRACT

As AI systems proliferate, their greenhouse gas emissions are an increasingly important concern for human societies. We analyze the emissions of several AI systems (ChatGPT, BLOOM, DALL-E2, Midjourney) relative to those of humans completing the same tasks. We find that an AI writing a page of text emits 130 to 1500 times less CO2e than a human doing so. Similarly, an AI creating an image emits 310 to 2900 times less. Emissions analysis do not account for social impacts such as professional displacement, legality, and rebound effects. In addition, AI is not a substitute for all human tasks. Nevertheless, at present, the use of AI holds the potential to carry out several major activities at much lower emission levels than can humans.

Motivation & Objective

• To quantify and compare the carbon emissions of AI systems versus humans in writing and illustrating tasks.
• To assess whether AI-based content creation offers lower environmental impact than human-driven processes.
• To evaluate the potential for AI to reduce greenhouse gas emissions in creative and writing workflows.
• To highlight the environmental benefits of AI while acknowledging limitations such as social impacts and task applicability.

Proposed method

• Measured carbon emissions for AI systems (ChatGPT, BLOOM, DALL-E2, Midjourney) during text generation and image creation tasks.
• Estimated human carbon emissions based on average energy use per task, including device operation and environmental context.
• Used standardized life cycle assessment methods to calculate CO2e emissions for both AI and human activities.
• Compared emissions per output unit (e.g., per page of text or per image) across multiple AI models and human performers.
• Accounted for training and inference phases in AI emissions, with inference being the primary focus for operational use.
• Applied conservative estimates for human energy use, including laptop and internet infrastructure contributions.

Experimental results

Research questions

• RQ1 How do the carbon emissions of AI-generated text compare to those of human-written text per page?
• RQ2 What is the ratio of carbon emissions between AI-generated images and human-created illustrations?
• RQ3 To what extent do AI systems reduce greenhouse gas emissions in creative and writing tasks compared to human labor?
• RQ4 What are the key factors influencing the environmental efficiency of AI versus human content creation?
• RQ5 How do emissions from AI inference compare to the cumulative emissions of human task performance?

Key findings

• AI writing a single page of text emits between 130 and 1500 times less CO2e than a human performing the same task.
• AI generating a single image emits between 310 and 2900 times less CO2e than a human creating the same image.
• The emissions advantage of AI is consistent across multiple models, including GPT-based and diffusion-based systems.
• The study identifies inference-phase emissions as the dominant factor in AI’s carbon footprint, with training emissions being relatively minor for repeated use.
• Even with conservative human energy use estimates, AI still shows a substantial emissions advantage in both writing and illustration tasks.
• The environmental benefits of AI are significant but do not account for social impacts such as job displacement or ethical concerns.