"AI perceptions of gender"
summary
Claims that bias can be shifted but not erased, demonstrated by training DCGAN and progressive GANs with 3D-rendered masculine feminine and non-binary faces, using transfer learning and “disruption” (injecting new data mid‑training); outputs shift toward fluidity but repeatedly revert to original masculine bias. A chained pipeline—Frankenstein’s telephone—links AttnGAN, DeepLab, SPADE/COCO and image→text and shows misgendering reemerge from caption training data. Built interactive mosaics and a color‑based multi‑label classifier that generates user RGB identities; it works as critique but requires intentional dataset bias to expose diversity.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:08AI artist and research focus
- 2:12AI interpretations of gender
- 3:38Set in stone debiasing GANs
- 8:58Gender as experience
- 9:40Frankenstein's telephone AI chain
- 15:23Generative AI biases research
- 17:53Gender tapestry color classifier
- 21:32Key findings and reflections
- 25:08Why this work matters
related talks
47mRoll for Initiative: how to make the world of AI a more ethical place
41mWhen AI Failures Reflect Our Own Social Biases
22mGender Bias in Artificial Intelligence
56mGlobal Bias in AI: When Western Data Shapes the World
23mDeveloping an intersectional framework to analyze biases in artificial intelligence and deep neural
11mAddressing Privacy, Bias, and Fairness in AI Systems
40mEthical AI: Addressing Bias & Algorithmic Fairness
47mEthical Machine Learning