Building fair, ethical, and responsible AI with the Responsible AI Toolkit
summary
Google’s Responsible AI Toolkit provides concrete tools and results: use Fairness Indicators to slice a restaurant-review sentiment model and find lower accuracy and notably higher false positive rates on reviews containing gendered terms (male reviews showing the biggest FPR gap). Use Language Interpretability Tool to edit 19 pronoun pairs, see short three-token reviews flip labels, highlight tokens like “waitress,” and run TCAV to show a female-gender concept increases negative sentiment score. LIT is released and framework-agnostic.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 1:19What is Responsible AI
- 2:01Building a Sentiment Classifier
- 3:25Using Fairness Indicators
- 6:27Language Interpretability
related talks
37mAn update on AI and FHIR
02mInterview with Milan Gospic
58mThe Modern Trolley Problem - Responsible AI Principles
39mHow to Use AI with Playwright for UI & API Testing
25mYou put what, where? What FHIR should be doing about AI/ML
41mDigital Dystopia or Artistic Utopia? The AI Creative Quandary
27mStereo-tAIping: Who Gets to Write the Story?
15mUsing software tools and AI to audit e-commerce food regulations