LLM agents show film narratives amplify gender gaps in attitude surveys
734 character agents from 160 films reveal narratives exaggerate gender differences more than real life.
A new arXiv paper proposes a framework that turns fictional film characters into 'surveyable LLM agents' to measure the gender values encoded in cultural narratives. The team—Vivienne Bihe Chi, Reyhan Jamalova, Lyle Ungar, and Sharath Chandra Guntuku—selected 160 U.S. films spanning 1990 to 2019 and built 734 character agents. They condensed personas from script dialogue and scene descriptions using expert-style reflections, then simulated responses to the World Values Survey on gender attitudes.
Without any explicit demographic prompting, the agents reproduced systematic gender differences—however, the gaps were exaggerated compared to real-world survey data. The agents also showed greater decade-to-decade volatility than actual human populations, suggesting that film narratives amplify rather than homogenize gender contrasts. This finding challenges the 'cultivation theory' notion that mass media mainstreams attitudes by providing consistent inputs. The authors argue that AI systems trained on film corpora may inherit this stylized polarization before any model-level amplification occurs.
- Built 734 LLM agents from 160 U.S. films (1990–2019) using script dialogue and scene descriptions.
- Agents reproduced gender attitude gaps without explicit demographic cues, but exaggerated them vs. real surveys.
- Gender differences were amplified by narrative, not smoothed—challenging cultivation theory's mainstreaming assumption.
Why It Matters
If AI learns from films, it may inherit exaggerated gender biases—long before any model-level reinforcement.