Deep learning holds promise for supporting radiologists by addressing challenges such as high workloads, increasing imaging volumes, and inconsistencies in image interpretation. However, current models require extensive annotations to work efficiently. The annotation of 3D medical images demands substantial time and expert effort, restricting the scalability of clinical AI applications. This work explores whether radiology reports, which are readily available and semantically rich, can serve as weak supervision for medical image segmentation. We investigate a contrastive vision-language model trained to align 3D computed tomography (CT) scans with free-text reports and probe the resulting representations using interpretability techniques. By analyzing attribution patterns extracted from the model, we assess whether it captures spatially meaningful signals despite lacking segmentation labels. This approach aims to reduce reliance on manual annotations and move toward scalable, label-efficient segmentation pipelines. The resulting code and comprehensive 3D visualizations can be found at https://github.com/injardav/CT-CLIP-UT .