Reconciling the contrasting narratives on the environmental impact of large language models

Shaolei Ren, Bill Tomlinson, Rebecca W. Black, Andrew W. TorranceScientific 2024 · DOI

This paper, published in Scientific Reports, tackles the polarizing debate over the environmental cost of Artificial Intelligence. While headlines often focus on the massive energy and water requirements of data centers, this study provides a necessary counter-perspective: how do those costs compare to the human labor they might replace or augment?

TL;DR

By applying Life Cycle Assessment (LCA) to the task of writing, researchers found that Large Language Models (LLMs) are 40 to 4400 times more resource-efficient than human workers in the U.S. across energy, carbon, water, and cost. Even in lower-footprint regions like India, LLMs remain significantly more efficient, though the gap narrows as model size increases.

What Problem Does This Paper Address?

The discourse around AI sustainability is currently split into two "contrasting narratives." One side views LLMs as a "sustainability problem" due to the massive resource consumption required for training and inference. The other sees them as a "sustainability solution" that can optimize industries and reduce human-driven environmental footprints.

The authors identify a critical gap: most studies look at AI's impact in a vacuum. To truly understand the environmental trade-offs, we must compare AI output to the "baseline" of human labor. If an AI uses 0.017 kWh to write a page that would have cost a human 0.85 kWh in residential electricity alone, the net environmental impact of the task actually decreases.

Methods

The researchers conducted a Life Cycle Assessment (LCA) focusing on a specific functional unit: writing one 500-word page of content.

  • AI Modeling: They analyzed two models: Llama-3-70B (representing a "typical" powerful model) and Gemma-2B-it (a "lightweight" model). They calculated operational energy (GPU + non-GPU server overhead) and used a conservative "factor of 2" multiplier to account for embodied "Scope 3" emissions (manufacturing hardware).
  • Human Modeling: This is the study's most novel contribution. They didn't just measure the calories a human burns while typing. Instead, they amortized a human’s total residential environmental footprint (electricity, water, carbon) over the time it takes to write a page (estimated at 1.67 hours).
  • Geographic Comparison: They compared impacts for residents of both the United States and India to account for different standards of living, wage levels, and grid carbon intensities.

Datasets

The study did not introduce new machine learning benchmarks but synthesized several high-quality public datasets for its LCA:

  • Energy/Carbon: U.S. Energy Information Administration (EIA) data and Climate Transparency reports for India.
  • Water: U.S. Geological Survey (USGS) and World Bank data on per-capita water withdrawal and consumptive rates.
  • Economic: U.S. federal minimum wage and India’s Chief Labour Commissioner guidelines.
  • Technical: Performance data for Nvidia H100 and A100 GPUs.

Key Insights and Ideas

  • The "Amortized Human" Concept: The authors argue that if a company "rents" 40 hours of a person's life per week, the environmental cost of that labor includes the infrastructure required to keep that person alive and functioning during those hours. This reframes human labor as a high-resource "industrial process."
  • The Efficiency Gap is Massive: In the U.S., the human-to-LLM ratio for carbon emissions is 53:1 for Llama-3-70B. For the lightweight Gemma model, it jumps to a staggering 4400:1.
  • Geography Dictates the Margin: Because the average U.S. resident has a much higher per-capita footprint than an Indian resident, the "savings" from switching to AI are much higher in the West. In India, a typical LLM is only about 3.4 times more energy-efficient than a human, suggesting that as models grow (like the 405B parameter versions), AI could eventually become less efficient than human labor in developing nations.
  • Model Size is the Deciding Factor: The move from a 70B model to a 2B model improved efficiency by roughly 100x, highlighting that "small AI" is the clear path for sustainability.

Key Results

The quantitative advantages of LLMs over human labor (U.S. baseline) for writing one page:

  • Energy: LLMs use 0.020 kWh (Llama) or 0.00024 kWh (Gemma) vs. 0.85 kWh for a human.
  • Carbon: LLMs emit 15g (Llama) or 0.18g (Gemma) vs. 800g for a human.
  • Water: LLMs consume 140ml (Llama) or 1.69ml (Gemma) vs. 5.7 liters for a human.
  • Cost: LLMs cost ~$0.08 (API equivalent) vs. $12.08 (U.S. min wage).

Limitations and Future Work

The authors are careful to note several caveats:

  • Quality vs. Quantity: The study measures resource utilization, not the quality or creativity of the text. It assumes 500 words of AI text is functionally equivalent to 500 words of human text, which is not always true.
  • The Rebound Effect (Jevons Paradox): Increased efficiency often leads to increased total consumption. If writing becomes "free" environmentally, we might simply produce 10,000x more text, potentially negating the savings.
  • Scaling Laws: As frontier models continue to grow in size, their energy demands may eventually outpace the "amortized" cost of a human, especially in regions with lower per-capita footprints.
  • Social and Ethical Costs: The paper acknowledges that environmental efficiency does not account for job displacement, the spread of misinformation, or the erosion of human creativity.

Summary updated

Abstract

The recent proliferation of large language models (LLMs) has led to divergent narratives about their environmental impacts. Some studies highlight the substantial carbon footprint of training and using LLMs, while others argue that LLMs can lead to more sustainable alternatives to current practices. We reconcile these narratives by presenting a comparative assessment of the environmental impact of LLMs vs. human labor, examining their relative efficiency across energy consumption, carbon emissions, water usage, and cost. Our findings reveal that, while LLMs have substantial environmental impacts, their relative impacts can be dramatically lower than human labor in the U.S. for the same output, with human-to-LLM ratios ranging from 40 to 150 for a typical LLM (Llama-3-70B) and from 1200 to 4400 for a lightweight LLM (Gemma-2B-it). While the human-to-LLM ratios are smaller with regard to human labor in India, these ratios are still between 3.4 and 16 for a typical LLM and between 130 and 1100 for a lightweight LLM. Despite the potential benefit of switching from humans to LLMs, economic factors may cause widespread adoption to lead to a new combination of human and LLM-driven work, rather than a simple substitution. Moreover, the growing size of LLMs may substantially increase their energy consumption and lower the human-to-LLM ratios, highlighting the need for further research to ensure the sustainability and efficiency of LLMs.
BibTeX
@inproceedings{ren2024reconciling,
  author = {Shaolei Ren and Bill Tomlinson and Rebecca W. Black and Andrew W. Torrance},
  title = {{Reconciling the contrasting narratives on the environmental impact of large language models}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-76682-6},
}
Figure