Designing levels—the design of the space in which a player moves, puzzles, fights, and explores—is one of the most labor-intensive and skill-intensive aspects of game development. Over the last decade, PCG (procedural content generation) has transitioned from being purely autonomous to mixed-initiative and human-AI co-creative, wherein the designer and an AI algorithm modify content in concert with one another versus the algorithm generating content fully autonomously. This paper collects together findings relating to human-AI co-creation from mixed-initiative PCG research, human-computer interaction studies, and recent reports investigating the use of generative AI within the game development industry. It classifies approaches to co-creation, presents example systems (Tanagra, Sentient Sketchbook, Ropossum, Lode Encoder, RL Brush, Baba Is Y'all, Promethean AI), and investigates implications for design, iteration speed, and production costs. The paper argues that co-creation does not inherently accelerate or improve design quality; rather, it depends on the degree to which the designer retains control and autonomy, the transparency of the algorithm's suggestions, and the framing of the designer as more than an editor of suggestions offered by an autonomous algorithm. This paper concludes with recommendations for practitioners aiming to incorporate co-creation into their design toolchains and future directions for research.