AI carbon footprint: is it really lower than human creative work?

A research study compared the CO2 emissions generated by AI versus humans for writing and illustration tasks. The result: AI produces up to 1,000 times fewer emissions than equivalent human activity. The finding is counterintuitive and important, but the study is clear about what it does not measure, including job displacement, legal questions around training data, and rebound effects.

What question does this research actually ask?

The study "The Carbon Emissions of Writing and Illustrating Are Lower for AI than for Humans" addresses a question that rarely comes up in AI debates: what is the actual environmental impact of using AI for creative tasks, compared to having humans perform those same tasks?

The answer challenges the common assumption. While training large language models is frequently criticized for its carbon footprint, the analysis shows that AI usage at the level of a single request is far less polluting than the equivalent human work.

What are the key numbers from the study?

AI-related emissions:

• Training GPT-3: approximately 552 metric tons of CO2e in total, amortized across billions of requests
• ChatGPT: approximately 2.2 grams of CO2e per request (training and operation included)
• BLOOM: approximately 1.6 grams of CO2e per request

Human-related emissions:

• Human writing: between 180 and 1,427 grams of CO2e per page produced
• Human illustration: between 690 and 5,500 grams of CO2e per image created

In concrete terms, producing one written page generates between 80 and 650 times more emissions when done by a human than when done by ChatGPT. For illustration, the gap can reach a factor of 1 to 2,500.

How did researchers measure these emissions?

The researchers conducted a quantitative analysis based on published data on the environmental impacts of AI systems and human activities. The method rests on three pillars:

• AI emissions calculation: training costs amortized across the total number of requests, plus operational emissions per request
• Human emissions estimation: average annual emissions per person, adapted to specific writing and illustration tasks
• Direct comparison: a ratio calculated between AI emissions and human emissions for the same tasks

What does the study leave out?

The researchers are explicit about the limits of their analysis. Several important dimensions are not included in the calculations:

• Potential job displacement from replacing human tasks with AI
• Legal questions around intellectual property and training data
• Rebound effects: if AI enables producing 10 times more content at the same cost, total emissions can increase even if the per-unit footprint is lower
• Incomplete substitutability: AI does not replace all human tasks, and collaboration between the two can be more advantageous than full replacement

FAQ

Isn't training AI models very polluting?

Yes, the initial training is energy-intensive. Training GPT-3 produced approximately 552 metric tons of CO2e. But that cost is amortized across billions of requests, which brings the per-request footprint down to approximately 2.2 grams of CO2e — well below the emissions a human generates for the same task.

Will AI replace human creatives for environmental reasons?

No. The study does not argue for full replacement. The researchers emphasize that AI cannot substitute all human tasks, and that collaboration between AI and humans is often more advantageous than pure replacement, both in terms of quality and sustainability.

What is the rebound effect in the context of AI?

The rebound effect refers to the phenomenon where an efficiency gain leads to an increase in overall consumption. If AI allows producing far more content at lower cost, the total carbon footprint of the sector can increase even if the per-unit footprint is smaller.

Do these results apply to all types of AI tasks?

No. The study focuses specifically on writing and illustration tasks. The conclusions cannot be generalized to all AI applications, particularly those requiring intensive computation or frequent retraining on new models.

Conclusion

This research sheds light on a dimension often missing from AI debates: its environmental impact compared to human activity. For writing and illustration tasks, using AI is today significantly less polluting than equivalent human work. That does not mean AI is without impact, but the sustainability debate benefits from precise data rather than assumptions.

The authors recommend a collaborative approach between AI and humans, maximizing the strengths of each. How this impact evolves will need to be monitored closely, particularly as technology and society continue to change.

AI Partners helps organizations adopt AI in a way that is performant, ethical, and responsible.