Traditional steganography is the practice of embedding a secret message into an image by modifying the information in the spatial or frequency domain of the cover image. Although this method has a large embedding capacity, it inevitably leaves traces of rewriting that can eventually be discovered by the enemy. The method of Steganography by Cover Synthesis (SCS) attempts to construct a natural stego image, so that the cover image is not modified; thus, it can overcome detection by a steganographic analyzer. Due to the difficulty in constructing natural stego images, the development of SCS is limited. In this paper, a novel generative SCS method based on a Generative Adversarial Network (GAN) for image steganography is proposed. In our method, we design a GAN model called Synthetic Semantics Stego Generative Adversarial Network (SSS-GAN) to generate stego images from secret messages. By establishing a mapping relationship between secret messages and semantic category information, category labels can generate pseudo-real images via the generative model. Then, the receiver can recognize the labels via the classifier network to restore the concealed information in communications. We trained the model on the MINIST, CIFAR-10, and CIFAR-100 image datasets. Experiments show the feasibility of this method. The security, capacity, and robustness of the method are analyzed.