Many empirical studies focus on socio-technical activity in social coding platforms such as GitHub, for example to study the onboarding, abandonment, productivity and collaboration among team members. Such studies face the difficulty that GitHub activity can also be generated automatically by bots of a different nature. It therefore becomes imperative to distinguish such bots from human users. We propose an automated approach to detect bots in GitHub pull request (PR) activity. Relying on the assumption that bots contain repetitive message patterns in their PR comments, we analyse the similarity between multiple messages from the same GitHub identity, using a clustering method that combines the Jaccard and Levenshtein distance. We empirically evaluate our approach by analysing 20,090 PR comments of 250 users and 42 bots in 1,262 GitHub repositories. Our results show that the method is able to clearly separate bots from human users.