Collaboration between humans and models can only be effective if both are aligned in their focus and understanding. In the past, research has mainly focused on correlation as a metric of human-model alignment. However, some results show that higher correlation does not lead to better task performance. We suspect a gap between shared attention and shared understanding. Attending to similar regions of an image does not necessarily mean that the same concepts are recognized there. Consequently, this thesis first examines whether human gaze and the model attention of a multimodal Transformer model attend to the same region in a VQA task and then investigates if they recognize the same concepts there. We find a correlation between human gaze and model attention, although higher correlated question-image pairs do not show higher VQA accuracy. Furthermore, we observe only a small overlap between model concepts and human annotated segmentation. A possible explanation is that shared attention is insufficient for human-model alignment and that concept alignment is necessary. However even that may not be enough as our results show that higher concept overlap also does not lead to significant improvement in VQA accuracy. Overall, this thesis demonstrates that shared attention is not shared understanding, and neither is sufficient for higher VQA accuracy.