Limits to AI Growth: The Ecological and Social Consequences of Scaling

TL;DR

This paper interrogates the feasibility and desirability of continuing to scale frontier AI given growing ecological and social costs. Using a social-ecological-technical (SET) lens and system dynamics, it maps four perspectives—technical, economic, ecological, and social—and employs archetypes like the limits-to-growth to model AI scaling dynamics. It identifies internal limits to scaling (performance, data, energy) and highlights externalized harms from unlimited capital expenditure and weak accountability, including labor exploitation and environmental injustice. The authors advocate frugal or sufficient AI and a reorientation of progress metrics to avoid overshoot and potential collapse, offering a pathway toward sustainable, mindful AI development.

Abstract

The accelerating development and deployment of AI technologies depend on the continued ability to scale their infrastructure. This has implied increasing amounts of monetary investment and natural resources. Frontier AI applications have thus resulted in rising financial, environmental, and social costs. While the factors that AI scaling depends on reach its limits, the push for its accelerated advancement and entrenchment continues. In this paper, we provide a holistic review of AI scaling using four lenses (technical, economic, ecological, and social) and review the relationships between these lenses to explore the dynamics of AI growth. We do so by drawing on system dynamics concepts including archetypes such as "limits to growth" to model the dynamic complexity of AI scaling and synthesize several perspectives. Our work maps out the entangled relationships between the technical, economic, ecological and social perspectives and the apparent limits to growth. The analysis explains how industry's responses to external limits enables continued (but temporary) scaling and how this benefits Big Tech while externalizing social and environmental damages. To avoid an "overshoot and collapse" trajectory, we advocate for realigning priorities and norms around scaling to prioritize sustainable and mindful advancements.

Figures (8)

• Figure 1: Limits to growth meadows_thinking_2011meadows_limits_1972 archetype senge_fifth_1997 template, adapted from sterman_business_2000. The archetype demonstrates that growth cannot continue forever because while the state of the system will grow at first, it will then slow until it reaches a state of equilibrium due to the limits dictated by the carrying capacity.
• Figure 2: Compute of ML models over time. Top: Select models from 1950-2024; Bottom left: Filtered for models in the pre deep learning era (1950-2010); Bottom right: Filtered for models in the the deep learning and large scale era (2010-2024). Data: epoch_ai_data_2024sevilla_compute_2022.
• Figure 3: Scaling of training data, investment, usage, energy, water, and carbon emissions of AI models. A: Language training dataset size over time, data: epoch_ai_data_2024sevilla_compute_2022. B: Private investment in AI, data: nestor_maslej_ai_2024. C: ChatGPT website visits (2022-24), data: Similarweb Similarweb. D: Maximum power consumption enabled by GPUs, data: kindig_ai_2024. E: Water usage for training models, data: li_making_2023shaolei_ren_how_2023pranshu_verma_bottle_2024barr_llamas_2023. F: Emissions and parameters for AI models, data: nestor_maslej_ai_2024 .
• Figure 4: The dynamics of AI scaling across perspectives (colour-coded) with social and environmental damages, showing that multiple reinforcing loops drive scaling while only one feedback loop limits scaling (i.e., the capacity of model scaling). The loops are R1: Economies of scale, R2: Competitive pressure loop, R3: AI hype and speculation, R4: Scaling driven by hype, R5: Scaling driven by capital investment, R6: Infrastructure development, B1: Capacity of model scaling.
• Figure 5: As models scale, the rise in hardware development implies increased energy and water usage which leads to efforts to make resource consumption more efficient. While this does lead to some efficiency gains for energy and water, there is also a rebound effect that contributes to further scaling. The loops are R1: AI model scaling, R2: Rebound effect of water and energy efficiency, R3: AI hardware scaling, B1: Water and energy efficiency gains, adapted from laurenti_unintended_2016.