“An increasing number of studies seek to assess the net environmental impact of artificial intelligence (AI) systems, weighing both positive and negative effects. This is a critical topic, as the net impact of AI is of great societal relevance yet challenging to determine. In this article, we review current methods for the assessment of direct and indirect carbon impacts of AI systems, including those that are transferred from the more general domain of information and communication technologies. We identify common principles that are shared across the majority of frameworks and the measurement challenges that arise specifically in the context of AI. We apply our findings to a previously published case study, demonstrating that refactoring a calculation to conform to the principles identified by established frameworks has a large impact on the result. We also quantify the sensitivity of the final estimate to key parameters used in the impact calculation. Carbon impact results prove highly sensitive to methodological choices, highlighting the need for more transparent, consistent, and AI-specific approaches. Today’s frameworks fail to capture AI’s distinctive characteristics, including its indirect effects, with sufficient accuracy to inform decision-making around AI’s environmental impact.”