Humor is widely considered a mechanism for fostering rapport and relational closeness in human–machine communication (HMC), yet which chatbot humor styles work—and why—remains poorly understood. Drawing on humor style theory and co-constructive perspectives on HMC, this study examines how affiliative, aggressive, self-defeating, and self-enhancing humor styles shape user evaluations. In a between-subjects experiment (N = 231), participants interacted with a chatbot employing one of four humor styles, rated on funniness, sociality, pleasantness, and popularity. Results show that humor effects depend more on interpretation than on design: affiliative readings enhanced evaluations across all styles, while misalignments proved consequential. Most strikingly, self-defeating humor received the highest evaluations, while self-enhancing humor underperformed most severely. This reflects an ontological asymmetry constitutive of HMC: whereas human–human communication rewards affiliation, HMC appears to reward hierarchy—the machine that jokes at its own expense affirms the proper status order; the machine that signals confidence transgresses it.