LLMs fail Bechdel test: Human scripts show less gender bias

A new paper from researchers at the University of Pennsylvania (Govindu et al., arXiv:2606.24022) investigates gender bias in screenplays generated by large language models (LLMs) using the Bechdel test—a measure of women's representation that requires at least two named women talking to each other about something other than a man. The team compared 768 human-written film scripts against screenplays generated by GPT-5, Gemini 3 Pro, and Claude Sonnet 4.5. They found that human-authored screenplays are significantly more likely to pass the Bechdel test, indicating that LLM-generated scripts underrepresent women in key narrative roles.

Beyond the traditional Bechdel test, the authors introduced social network analysis metrics—such as centrality (character importance), homophily (same-gender clustering), and triadic relationships—to capture deeper representational biases. Under some of these measures, LLM scripts showed slightly less bias than human scripts (e.g., less extreme gender homophily), but all script types exhibited significant representational disparities. The study underscores that current state-of-the-art LLMs still perpetuate harmful gender stereotypes, and calls for richer quantitative auditing frameworks as AI becomes more involved in media production like filmmaking and journalism.