Scott Logic: Greener AI – what matters, what helps, and what we still do not know

Sep 17, 2025

—

Source URL: https://blog.scottlogic.com/2025/09/16/greener-ai-lit-review.html
Source: Scott Logic
Title: Greener AI – what matters, what helps, and what we still do not know

Feedly Summary: We recently undertook a literature review about the environmental impact of AI, across carbon, energy, and water. It offers practical strategies for teams to reduce impact today, while highlighting the gaps in measurement, reporting, and governance that still need to be addressed.

AI Summary and Description: Yes

Summary: The text explores the environmental impacts of artificial intelligence (AI), particularly focusing on large language models (LLMs), through a literature review aimed at understanding emissions throughout the AI lifecycle. It emphasizes the importance of measuring and managing AI’s environmental footprint, especially as the technology becomes more prevalent. The findings indicate a complex interplay between training and inference emissions, and highlight practical strategies to mitigate impacts.

Detailed Description:
The review presented in the text identifies critical environmental considerations that arise from the transition of AI from research to widespread use, particularly the significant emissions associated with LLMs. Key points include:

– **Environmental Impact Measurement**: The study investigates what metrics researchers are currently employing to gauge the environmental footprint of AI technologies:
– Assessing emissions throughout the AI lifecycle, including hardware production, training, inference, and operational use.
– Evaluating the methodologies used to capture resource consumption effectively.

– **Energy and Emissions Context**: It highlights the clear distinction between operational emissions (from running the AI) and embodied emissions (from hardware manufacturing). The review found that inference emissions could potentially exceed training emissions at scale, suggesting that operational efficiencies are vital for long-term sustainability.

– **Measurement Tools and Strategies**:
– The literature points to tools like Carbontracker and CodeCarbon for accurately tracking energy and emissions.
– Various strategies are explored for reducing the environmental impact of AI, such as:
– Optimizing AI models to be smaller and more efficient.
– Energy-aware training that considers the timing of resource consumption.
– Implementing life cycle approaches for hardware procurement, extending its usable life.

– **Challenges and Barriers**: The text outlines significant obstacles like inconsistent reporting of emissions data, lack of accessibility to hardware emissions data, and the potential for demand-side governance to inadvertently negate operational efficiencies.

– **Governance and Standards**: The necessity for established frameworks is underscored, including standardized reporting to enable comparability across studies and clear definitions of boundaries for lifecycle assessments. It emphasizes that comprehensive measures, transparent reporting, and system-level governance are crucial for effectively mitigating AI’s environmental impact.

Key Takeaways:
– The distinction between training and inference emissions is critical for understanding AI’s lifecycle impact.
– There are practical, actionable strategies that can be adopted today to enhance AI’s sustainability.
– Without shared standards, the evaluation of AI’s environmental performance remains challenging, risking superficial gains in efficiency metrics.
– Fostering transparency in reporting and governance structures is essential for meaningful progress toward integrating sustainability in AI development.

Actionable Recommendation: Security and compliance professionals need to ensure that their organizations adopt frameworks that allow for transparent carbon tracking and reporting, which can enhance both corporate responsibility and competitive advantage in developing AI technologies.

1 2 2025 5 a access accessibility Act age AGI AI AI development ai model AI models AI technologies All allow and app Arch art artificial Artificial Intelligence as assessment assessments at ated aware Bi bot C carbon carbon tracking challenge challenges CI CIA CleaR co code competitive competitive advantage compliance compliance professionals consumption Context core corporate responsibility CoT critical cross Current D data day de DeFi definition definitions demand development e effective efficiency efficient emissions end energy environment environmental environmental considerations environmental impact ERP evaluation exp fact for framework frameworks g Gen Go governance governance structure governance structures gs H hardware hardware production high Highlight HR http HTTPS impact in Inference Intel intelligence inter io Iron ite k Key l language language model language models large large language model large language models Large Language Models (LLMs) led level Li life lifecycle assessment line Lite llm llms lm logic long low M man manufacturing matt mean measures methodologies metrics mission ML Mode model models N no o of off on ons operation operational operational efficiencies OPM opt organization organizations oS oss out over per performance play point porting potential pre pro procurement product production professionals Progress ps R rack rate RCE re red report reporting research researchers resource resource consumption responsibility review Risk Ro s Sable Scale search sec security security and compliance SHA side Sig size sizes small SoC source SSE SSO standards STIG strategies structures study sustainability system T team Teams tech technologies technology ted text the to tool tools TP tracker tracking training training and inference transition transparency transparent under up US use V val Valuation Vantage Ware Wi x z