Megha @meghag.bsky.social · 25/06/2026A huge shoutout to the co-author team 😊: @stephanietwang.bsky.social, @friedler.net, and Danaé Metaxa. Read our full paper here: arxiv.org/abs/2606.24022 000
Megha @meghag.bsky.social · 25/06/2026Our work supports the use of SNA in LLM bias audits, and the broader adoption of alternative evaluation strategies. However, even if LLM outputs are “unbiased” and equitable, a core issue remains: AI is trained—without consent—on the work of human artists it attempts to replace. 100
Megha @meghag.bsky.social · 25/06/2026We find that human-written scripts perform better than LLMs on the Bechdel Test. However, LLMs like GPT outperform human-written scripts on other social network analysis (SNA) measures. Above all, whether they are produced by LLMs or human-written scripts, women remain poorly represented. 100
Megha @meghag.bsky.social · 25/06/2026We developed a dataset of human-written screenplays and counterpart LLM screenplays from 3 models – GPT, Gemini, and Claude. We then analyzed Bechdel Test performance and the underlying social network of each film (rather than the raw text) to compare gender bias in human-written and LLM scripts. 100
Megha @meghag.bsky.social · 25/06/2026Films both reflect and construct cultural narratives. Separately, LLMs exhibit documented biases. With 1) the increased presence of LLMs in film and 2) how the film medium lends itself to examining representational bias, we select this context to audit for bias in LLMs relative to human outputs. 100
Megha @meghag.bsky.social · 25/06/2026Now that AI is writing our movies, will it let women talk? In our paper, we leverage the #BechdelTest and other social network-based measures to evaluate LLM-generated movie scripts for gender bias. Join us today at ACM #FAccT2026: Thursday, June 25th at 10:45 am in Joyce to hear more!brookings.eduHollywood writers went on strike to protect their livelihoods from generative AI. Their remarkable victory matters for all workers. | Brookings 141