<p>The main objective of this review is to summarise the evidence about the artificial intelligence models in the assessment of shoulder magnetic resonance imaging.</p><p>This scoping review included observational studies focused on the use of artificial intelligence to improve the diagnosis of shoulder injuries in magnetic resonance imaging. The studies that mixed joints or whose approach was theoretical were excluded. Regarding the analysis of each study, the data included were study type, population profile, sample size, method of deep learning, outcome variables, accuracy measurements, and statistical significance.</p><p>26 studies with a total of 9019 subjects were included. Seven studies focused on the accuracy of segmentation; thirteen, on the diagnosis, five, on the image quality, and one, on the acquisition time. For these purposes, all studies bet on convolutional neural networks, with seven choosing the 3D U-Net model and seven opting for more specific models such as 1-Lipschitz neural network, V-Net, ViVGG19, or CapsNet. Furthermore, six studies opted for the conjunction of models such as block-based AlexNet, GoogLeNet inception v3, ResNet or SqueezeNet with the 3D U-Net model. These models obtained statistically significant improvements in image quality (<i>p</i> = 0.001–0.028) and accuracy of segmentation (<i>p</i> = 0.001–0.043). However, despite obtaining a slight optimization of diagnostic sensitivity, these differences were not significant in all studies (<i>p</i> = 0.001–0.99).</p><p>Deep learning stands out as a promising tool in the management of imaging tests in shoulder pathology, but more research is required to draw consistent conclusions.</p>