How to Build a Gaslighting Detector: A Multivariate Pattern of Emotion Salience Marks Gaslighting in Natural Language

Gaslighting is a form of psychological manipulation involving denial or distortion of another person’s perspective and/or perception of events. Gaslighting within interpersonal interactions is often a subtle and/or covert phenomenon that occurs under the guise of emotional support. A person who is skilled in gaslighting might mask their intent by avoiding particular ‘red flag’ language constructions (e.g., words, phrases). This variability in lexical use presents a significant challenge for classification methods based on detecting canonical distributions of words. We describe an alternative method and open-source R software application (Gaslighting) for classifying gaslighting in language samples of any length. Young adults (N=1,034) generated formal definitions of and stories about gaslighting. We indexed the lexical content of these narratives to derive a 22-dimension affective signature for gaslighting. This composite emotion vector yields a quantitative base for computing likelihood estimates of gaslighting in any new language sample. We derived expected distributions of gaslighting by computing distance metrics between emotion vectors for a training set of gaslighting stories (n=400) relative to definitions of gaslighting (N=1,017). We then examined classification accuracy using signal detection metrics for a holdout set of gaslighting stories (N=619) relative to expository language samples extracted from eight Wikipedia keyword searches (e.g., birds, furniture, art) (N=9600). Classification for gaslighting vs. non-gaslighting texts was high (85%) with a corresponding sensitivity of 72.6% and specificity of 78.1%. We discuss advantages of this method of emotion pattern matching in naturalistic language samples, including applications to other affective states or traits. We further outline computational assumptions, limitations, and caveats for using the algorithm on untrained/untested language samples.