In high-risk areas such as medicine, neural networks have established themselves as extremely powerful tools - especially for segmenting endoscopic image data. Despite their high accuracy, however, they still pose a key challenge: As "black boxes", their decision-making processes are difficult to understand and they output predictions even when inputs lie outside the learned data space. Systems that do not communicate their boundaries transparently and whose predictions are potentially erroneous therefore pose a hidden and significant risk. For this reason, the mere performance of an AI system is not enough for safe and effective integration into everyday clinical practice - key mechanisms need to be developed that make AI systems aware of their own limitations and thus create the reliability and trust that is essential for safe application in healthcare.