Clickbait is an attractive yet misleading headline that lures readers to commit click-conversion. Development of robust clickbait detection models has been, however, hampered due to the shortage of high-quality labeled training samples. To overcome this challenge, we investigate how to exploit human-written and machine-generated synthetic clickbaits. We first ask crowdworkers and journalism students to generate clickbaity news headlines. Second, we utilize deep generative models to generate clickbaity headlines. Through empirical evaluations, we demonstrate that synthetic clickbaits by human entities and deep generative models are consistently useful in improving the accuracy of various prediction models, by as much as 14.5% in AUC, across two real datasets and different types of algorithms. Especially, we observe an improvement in accuracy, up to 8.5% in AUC, even for top-ranked clickbait detectors from Clickbait Challenge 2017. Our study proposes a novel direction to address the shortage of labeled training data, one of fundamental bottlenecks in supervised learning, by means of synthetic training data with reinforced domain knowledge. It also provides a solution for distinguishing between bot-generated and human-written clickbaits, thus aiding the work of moderators and better alerting news consumers.