In this paper, we address the following question: what is the difference between predicates like eat and call, on the one hand, and offend and break, on the other hand, that is responsible for the presence/absence of weakness of will inference? The answer we arrive at is that the former, but not the latter, express actions that can be fully controlled by the agent. We argue that weakness of will can be used as a linguistic test to distinguish intentional versus accidental actions. We will see that the distinction between intentional and accidental actions cannot be detected by other linguistic tests, such as modification by (un)intentionally or by rationale clauses introduced by in order to (e.g., Jackendoff 1972, 1995; Farkas 1988, 1992). The intentional/accidental distinction is crucial for analysing many phenomena that have recently received a significant amount of attention in the literature, such as generalized subject obviation (Oikonomou 2016; Stegovec 2017, 2019; Kaufmann 2019; Szabolcsi 2010, to appear, a.o.), licensing of polarity sensitive items (Szabolcsi 2010; Zu 2018; Goncharov 2020a, a.o.), and aspect choice in Slavic (Despi ́c 2020; Goncharov 2020b, a.o.).