Background: Mixed audience reactions to movies can provide insights into societal trends and viewer types, yet predictors of such polarization remain understudied. This study presents an automated approach using a Large Language Model (LLM) to tag large-scale movie data, offering a more efficient alternative to manual coding or traditional natural language processing tools. The measure of audience polarization was derived from user ratings from Letterboxd.com. 

Objective: The current study examined whether (i) protagonists from marginalized groups (women, LGBTQ+, and non-white), (ii) release year, (iii) moral themes, and (iv) an experimental-style would predict audience polarization.

 Method: Films were randomly sampled from The Movie Database (TMDb). Metadata, including ratings, was scraped from Letterboxd.com, and the AI agent was then prompted with structured instructions to generate annotations for predefined movie characteristics, enabling large-scale automated tagging for analysis. 

Results: Four hypotheses were tested using separate linear regression models. 

Discussion: Moral themes were positively associated with audience polarization, whereas release year was negatively correlated. Other predictors were non-significant. However, interaction effects between marginalized group categories revealed interesting patterns. Conclusions are discussed in relation to prior research/theory, and study design limitations.