Evaluating the Environmental Impact of Generative AI in Healthcare: A Systematic Review of Current Practice

Background: Generative artificial intelligence (GenAI) is resource-intensive, raising concerns about its environmental sustainability in healthcare. This systematic review assessed current practices for measuring and reporting the environmental impact of GenAI applications in healthcare. <div> <br> </div> <div> Methods: We searched MEDLINE, Embase, preprint servers (ArXiv, BioRxiv, MedRxiv), and AI conference proceedings (NeurIPS, ICML, ICLR, IEEE) for primary research describing GenAI use in healthcare, published from 1st June 2018 to 10th July 2025. Studies were included if they provided quantitative measurement or qualitative discussion of GenAI's environmental impacts. Where GenAI was compared quantitatively with a non-GenAI solution, the relative environmental impact was evaluated, using the Mann-Whitney U test to assess for statistically significant differences where multiple comparisons were available. The protocol was registered prospectively on PROSPERO (CRD420251089924). </div> <div> <br> </div> <div> Findings: Of 1,513 records identified, 15 studies were included. Eight (53·3%) reported quantitative outcomes and all 15 included qualitative discussion. Quantitative measures comprised training and/or inference time (four studies), computational complexity (three studies), and direct energy usage and carbon dioxide emissions (two studies each). Direct GenAI versus non-GenAI comparisons were performed in four studies; in all, the environmental impact was greater for GenAI without a corresponding gain in task performance (p&lt;0.05 for the environmental impact measure in all comparisons). Qualitative discussion mostly addressed resource implications of model size or offered only generic acknowledgement of environmental impacts. No study assessed downstream effects (e.g. reductions in human labour or travel). </div> <div> <br> </div> <div> Interpretation: Few healthcare studies report or discuss the environmental impacts of GenAI, and those that do use heterogenous metrics. Although few head-to-head comparisons limit generalisability, available evidence suggests non-GenAI solutions can achieve comparable performance with lower resource requirements, highlighting the importance of considering environmental impact when selecting AI methods. Clearer guidance and standardised reporting frameworks are needed to improve transparency in healthcare GenAI. </div>