Conditional constructions (if–then) enable us to express our thoughts about possible states of the world, and they form an important ingredient for our reasoning and argumentative capabilities. Different types and argumentative uses have been distinguished in the literature, but their applicability to actual language use is rarely evaluated. This paper focuses on the reliability of applying classifications of connections between antecedents and consequents of conditionals to discourse, and three issues are identified. First, different accounts produce incompatible results when applied to (...) language data. Second, a discrepancy between theory and data was observed in previous studies, which sometimes discard existing classifications for being detached from actual language use. Finally, language users construct various cognitive relations between clauses of conditionals without being able to rely on overt linguistic features, which poses problems for the annotation of conditionals in argumentation and discourse. This paper addresses these issues by means of comparing theoretical types and actual uses of conditionals, by inspecting the dispersion of types in natural-language corpora, and by conducting an experiment in which the inter-rater reliability of classifications was assessed. The results show that the reliability of classifications of conditionals when applied to language data is low. With respect to the aforementioned issues, different classifications produced incompatible results, a discrepancy between theory and data was indeed observed, and low reliability scores indicated a largely interpretative nature of types of conditionals. Given these results, suggestions for the enhancement of reliability in corpus studies of conditionals and beyond are provided to enhance future classification design. (shrink)