We introduce a taxonomy of artificial intelligence (AI) experiments. Our taxonomy produces four types of AI experiments: conceptual AI experiments, stylized AI experiments, quasi-natural AI experiments, and natural AI experiments. At the core of our taxonomy is the sophistication of AI used, which we evaluate using a simple and robust proxy test of whether AI is developed exclusively for a research study. We discuss the advantages, disadvantages, and best use cases for each type and illustrate the use of each type in various examples. We provide a guide on how to choose the type of AI experiment that best fits a given research question.