While string obfuscation is a common technique used by mobile developers to prevent reverse engineering of their apps, malware authors also often employ it to, for example, avoid detection by signature-based antivirus products. For this reason, robust techniques for detecting obfuscated strings in apps are an important step towards more effective means of combating obfuscated malware. In this paper, we discuss and empirically characterize four significant limitations of existing machine-learning approaches to string obfuscation detection, and propose a novel method to address these limitations. The key insight of our method is that discriminative classification methods, which try to fit a decision boundary based on a set of positive and negative samples, are inherently bound to generalize poorly when used for string obfuscation detection. Since many different string obfuscation techniques exist, both in the form of commercial tools and as custom implementations, it is close to impossible to construct a training set that is representative of all possible obfuscations. We instead propose a generative approach based on the Naive Bayes method. We first model the distribution of natural-language strings, using a large corpus of strings from 235 languages, and then base our classification on a measure of the confidence with which a language can be assigned to a string. Crucially, this allows us to completely eliminate the need for obfuscated training samples. In our experiments, this new method significantly outperformed both an n-gram based random forest classifier and an entropy-based classifier, in terms of accuracy and generalizability.