The filter bubble is an intermediate structure to provoke polarization and echo chamber in social network, and it has become one of the most urgent issues for social media of the time. Previous studies usually equated filter bubbles with community structures and emphasized this exogenous isolation effect, but there is a lack of full discussion of the internal organization of filter bubbles. Here, we design an experiment for analyzing filter bubbles taking advantage of social bots. We deployed 128 bots to Weibo (the largest microblogging network in China), and each bot consumed a specific topic (entertainment or sci-tech) and ran for at least two months. In total, we recorded about 1.3 million messages exposed to these bots and their social networks. By analyzing the text received by the bots and motifs in their social networks, we found that a filter bubble is not only a dense community of users with the same preferences but also presents an endogenetic unidirectional star-like structure. The structure could spontaneously exclude non-preferred information and cause polarization. Moreover, our work proved that the felicitous use of artificial intelligence technology could provide an useful experimental approach that combines privacy protection and controllability in studying social media.