Transparent Screening for LLM Inference and Training Impacts

Researchers Arnault Pachot and Thierry Petit have introduced a novel framework for estimating the environmental footprint of large language models (LLMs) like GPT-4, Claude, and Llama 3. Published on arXiv (paper 2604.19757), the 'Transparent Screening' method addresses a critical industry blind spot: the opaque and often undisclosed energy and carbon costs associated with training and running massive AI models. Rather than attempting impossible direct measurements of proprietary systems, the framework provides an auditable proxy methodology. It converts natural-language descriptions of an intended AI application—such as 'customer support chatbot handling 10,000 queries per day'—into bounded environmental impact estimates, creating a reproducible and comparable baseline.

The core innovation is its focus on comparability and transparency under conditions of limited observability. The framework supports the creation of an online observatory where different models' estimated impacts can be compared side-by-side. This is crucial for enterprise decision-makers and sustainability officers who need to choose between vendors like OpenAI, Anthropic, or open-source options but currently lack standardized data. By linking estimates to their source methodologies, the tool aims to move the industry beyond vague claims toward accountable, science-based comparisons of AI's carbon cost, fostering more sustainable development practices.