Pronouns in Disguise: Anaphoric Variability, Structural Complexity, and Social Meaning
This talk investigates expressions that look like referential definite descriptions but behave like pronouns: epithets such as the idiot in English, and the Brazilian Portuguese 1PL form a gente (lit. 'the people', meaning 'we'). Despite their superficial form, these expressions exhibit the core properties of pronouns: e.g. they allow deictic/anaphoric readings, are exempt from Principle C effects, and lack descriptive content beyond ϕ-features. At the same time, they are distributionally and situationally more restricted than their plain pronominal counterparts.
Contrary to previous approaches (e.g. the imposter analysis of Collins & Postal 2012, or the DP-internal expletive hypothesis of Saab 2022), I argue that these “disguised pronouns” do not require any syntactic structure that is not motivated elsewhere: epithets have the same D-NP syntax as ordinary definites. They differ only in having a pronominal semantics, arising from an indexed D (characteristic of anaphoric definites, Schwarz 2013) and a dimension-shifted NP (also required for Donnellan-style referential readings, Varaschin & Machicao y Priemer t.a.). A gente has the same structure as the standard 1PL pronoun nós, differing only in bearing an additional set of 3SG CONCORD features that trigger hybrid agreement patterns (Smith 2021, Varaschin & Machicao y Priemer 2025, a.o.).
In both cases, the “disguised pronoun” is featurally more complex than its equivalent bare counterpart. Assuming a multidimensional semantics that distinguishes truth- from use-conditional meaning (Potts 2005, Gutzmann 2015), I propose that this complexity forces use-conditional content via Gricean Manner reasoning: complex forms survive because their additional features require them to grammaticalize a particular social meaning (a derogatory stance for epithets, informality for a gente), restricting their distribution to specific registers.
I close by showing that this hypothesis linking complexity to register predicts a gap in current language models: because LMs flatten register-conditioned variation during pretraining and alignment (Zanotto & Aroyehun 2024, a.o.), they underrepresent featurally complex forms like epithets and a gente, whose existence depends on use-conditional social meaning. I therefore suggest that a formally precise integration of syntax with semantics and (socio-)pragmatics offers a path toward more adequate models of language, both human and artificial.