Materials whose properties are determined mostly by their internal architecture—so-called metamaterials—have emerged as a growing field of study over the past decades. While the mechanical properties of these materials can be predicted using numerical methods, many applications also require the identification of architectures with specific target properties. Recently, a variety of generative machine learning approaches have been successfully applied to solve this inverse problem for a wide range of metamaterial designs. However, study after study focuses on the implementation of generative models on metamaterials with “nice“ parametrizations, while metamaterials with more intricate relationships between geometrical elements are almost entirely overlooked. Simultaneously, the true strength of advanced generative models emerges when intricate restrictions on the design space need to be learned. To demonstrate the existence of “survivorship bias“ in current literature and pinpoint its underlying cause, we examine a well-established class of kirigami metamaterials where dependencies between cuts yield complex design restrictions. When intersections between cuts in kirigami metamaterials are prohibited, the uniform Euclidean distance cannot be employed to measure similarities between such kirigami structures. We assess the capability of the four most popular generative design algorithms—the Variational Autoencoder (VAE), the Generative Adversarial Network (GAN), the Wasserstein GAN (WGAN), and the Denoising Diffusion Probabilistic Model (DDPM)—to generate such kirigami structures. We demonstrate that the reliance on this similarity metric is what hinders VAE and WGAN from learning to avoid intersections. Furthermore, we identify the reliance on Euclidean distance as one of the contributing factors to “survivorship bias“ and find that the investigated kirigami metamaterials might serve as a valuable benchmark for the future development of algorithms for designing mechanical metamaterials.