One of the primary goals of the visual system is to extract statistical regularities from the environment to build a robust representation of the world. Recent research on visual statistical learning (VSL) has demonstrated that human observers can implicitly extract joint probabilities between objects during streams of visual stimuli. In the real world, temporal predictability between scenes and places exists at both exemplar and categorical levels: whatever office you are in, the probability that you will step out in a zoo is much lower than the probability that you will enter a corridor. In a series of experiments, we tested to what extent people are sensitive to the learning of categorical temporal regularities based on the gist or semantic understanding of natural scenes. Our results suggest that the gist of a scene is automatically and implicitly extracted even when it is not task-relevant, and that implicit statistical learning can occur at a level as abstract as the conceptual gist representation.