Roll for Initiative: how to make the world of AI a more ethical place
summary
PhD work shows how to de-bias a GAN by altering training data and retraining, and the speaker reports using generated 3D humans and StyleGAN instead of real photos to avoid privacy harms. Cites MIT finding label errors up to 3.4% in popular image datasets, warns about stochastic parrots (large LMs) and biased outputs, and gives concrete mitigations: diversify teams, curate smaller datasets, tune sample weights and hyperparameters, audit with groups like the Algorithmic Justice League, and prototype a D&D module and insult-generator model.
AI-written summary · about
chapters (6)
- 0:00LINUX
- 3:58Campaign against Al Bias
- 16:27Prevention is better than cure
- 28:56Questions to ask
- 34:41Ways to stop bias
- 40:59What if it all goes wrong?
related talks
43mAI Engineer Ethics: Shaping Responsible Generative AI
47mEthical Machine Learning
41mWhen AI Failures Reflect Our Own Social Biases
19mMachine Intelligence With Humans in Mind
30mHow to Lie with AI: Understanding Bias, Ethics, and the Hidden Risks in ML
27m"AI perceptions of gender"
26m(Un)ethical Artificial Intelligence: How to Keep the AI Fair for Everyone
34mEthical framework for AI : Creating AI powered products that empower people