Socially responsible AI: Practical tools for robust design
summary
Argues responsible AI is just robust engineering, not vague ethics, and gives concrete tactics: use multiple and guard metrics and subgroup analysis to avoid Goodhart failures (example: a push-notification campaign gamed monthly active users, earned bonuses, then lost users); prefer interpretable models when performance matches (CORALS equals complex model); mitigate generative-AI bias by understanding downstream use (images of “lawyers” were whitened then patched by appending descriptors); recommends the Responsible Tech Playbook, an AI design alignment lens, and transparency that is legible auditable actionable.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 1:15What is responsible AI
- 5:00Technology shapes society
- 5:57Technology shapes our societies
- 8:28How to do responsible tech
- 10:25Thoughtworks and responsible tech
- 15:25Example
- 17:51Two kinds of environments
- 20:10Discrimination
- 21:40generative AI
- 24:38equality
- 24:51measuring equality
- 26:29metrics monitoring
- 27:04early warning systems
- 27:35sub populations
- 28:07manipulation
- 31:54transparency
- 33:46design thinking
- 36:53practical tools
- 38:55AI governance
- 39:23Recap
related talks
34mResponsible AI: The Key to Ethical, Safe and Inclusive Software Development
33mResponsible AI in Practice: Real-World Examples and Challenges
44mPanel: Building responsible, human-centric AI models
15mResponsible AI 2.0 : NO AI without RAI
35mResponsible AI: Ethics in Software Development
38mBuilding Responsible AI in the Age of GPT
01h 00mAI for Social Responsibility: Embedding (...) into AI
27mWriting the Playbook for Fair & Ethical Artificial Intelligence & Machine Learning