No abstract

Confession evidence is powerfully persuasive, and yet many wrongful convictions involving false confessions have surfaced in recent years (Innocence Project, 2021; National Registry of Exonerations, 2021). Although police are trained to corroborate admissions of guilt, research shows that most false confessions contain accurate details and other content cues suggesting credibility as well as extrinsic evidence of guilt. Hence, a method is needed to help distinguish true and false confessions. In this study, we utilized a corpus-based approach to outline the linguistic features of two sets of confessions: those that are presumed true (n = 98) and those that have been proven false (n = 37). After analyzing the two corpora in LIWC (Linguistic Inquiry and Word Count) to identify significant categories, we created a logistic regression model that distinguished the two corpora based on three identified predictors: personal pronouns, impersonal pronouns, and conjunctions. In a first sample comprised of 25 statements per set, the model correctly categorized 37 out of 50 confessions (74%); in a second out-of-model sample, the predictors accurately classified 20 of 24 confessions (83.3%). A high frequency of impersonal pronouns was associated with confessions proven false, while a high frequency of conjunctions and personal pronouns were associated with confessions presumed to be true. Several patterns were observed in the corpora. In the latter set of confessions, for example, “I” was often followed by a lexical verb, a pattern less frequent in false confessions. Although these data are preliminary and not to be used for practical diagnostic purposes, the findings suggest that additional research is warranted.