Ambiguous Label Learning (ALL), as an emerging paradigm of weakly supervised learning, aims to induce the prediction model from training datasets with ambiguous supervision, where, specifically, each training instance is annotated with a set of candidate labels but only one is valid. To handle this task, the existing shallow methods mainly disambiguate the candidate labels by leveraging various regularization techniques. Inspired by the great success of deep generative adversarial networks, we apply it to perform effective candidate label disambiguation from a new instance-pivoted perspective. Specifically, for each ALL instance, we recombine its feature representation with each of candidate labels to generate a set of candidate instances, where only one is real and all others are fake. We formulate a unified adversarial objective with respect to three players, i.e., a discriminator, a generator, and a classifier. The discriminator is used to detect the fake candidate instances, so that the classifier can be trained without them. With this insight, we develop a novel ALL method, namely Adversarial Ambiguous Label Learning with Candidate Instance Detection (A2L2CID). Theoretically, we analyze that there is a global equilibrium point between the three players. Empirically, extensive experimental results indicate that A2L2CID outperforms the state-of-the-art ALL methods.