In this paper, I criticize an argument presented by Dan Harris (forthcoming), but show how its conclusion could be secured by other means. On Harris’ view, if linguistics is a cognitive science, then its components, including semantics, must be modules—informationally encapsulated input systems (Fodor 1983). Harris then argues that a semantic module must output not a proposition, but a constraint on a class of propositions (cf. Carston 2012). The route from sentence meaning to speaker meaning is then a matter of satisfying these constraints. However, Harris’ argument neglects an alternative possibility, compatible with his claim that semantics is encapsulated: that the route to speaker meaning is not constraint satisfaction but development or modification of a highly general proposition, the latter being the output of semantics (Borg 2004). Furthermore, the premise that semantics is an encapsulated system must be defended in light of psycholinguistic data suggesting otherwise. These data present a substantial challenge for the view that semantics is informationally encapsulated, analogous to the challenges regarding the encapsulation of early vision (Ogilvie and Carruthers 2016). However, as I argue, a closer look at these data suggests that a content-based account of semantics is less plausible than the constraint-based account. In other words, there are empirical considerations that suggest that semantics must output constraints, not contents. These considerations are not decisive, but they suggest a non-question-begging way of defending a constraint-based view of semantics, as I show.