The Hidden Environmental Cost of Poor Coding Practices in TensorFlow and Keras Applications: A Study on Resource Leaks and Carbon Emissions

Poor coding practices in TensorFlow and Keras applications can sharply increase energy consumption and carbon emissions, according to a new study that quantifies the environmental toll of common resource leaks in machine learning code [1]. The study, posted to arXiv on June 18, 2026, examined two resource-leak smells: Improper Model Reuse (IMR) and Unreleased Tensor References (UTR). Controlled experiments showed IMR increased electricity consumption by approximately 32% and UTR by approximately 46%, with proportional rises in CO2 emissions [1][2]. Paired statistical tests indicated the differences were systematic and statistically significant [2]. Resource leaks in ML code can introduce hidden inefficiencies that elevate energy consumption and CO2 emissions, yet empirical evidence quantifying their environmental impact has been limited [2]. The authors wrote that the findings provide initial empirical evidence that resource-leak smells may degrade ML energy efficiency and environmental sustainability [2]. The paper recommends integrating resource-lifecycle management and energy-efficiency considerations into ML development [1][2]. The findings arrive as the computational demands of large language models draw increased scrutiny. DeepSeek, a Chinese AI company, reported training its V3 model for US$6 million, using approximately one-tenth the computing power consumed by Meta's comparable Llama 3.1 model [6]. Large language models are trained with self-supervised learning on vast amounts of text and contain many parameters [7]. The arXiv paper is accessible through the platform's Labs framework, which allows collaborators to develop and share new features directly on the site [1]. Hugging Face has collaborated with arXiv to embed interactive demos alongside papers, enabling users to try state-of-the-art research without writing code [3][4][5]. The integration, launched in 2022, has made over 12,000 open-source machine learning demos available through Hugging Face Spaces [3]. Researchers and community members can link demos to arXiv papers by including a paper citation in a Space's README file or by associating a model on the Hugging Face Hub with a Space [5]. The demo tab appears on arXiv abstract pages for papers in computer science, statistics, and electrical engineering and systems science categories [4].