KEYNOTE: Ethics is Not a Feature
summary
Claim: ethics cannot be tacked on as a feature but must shape data, modeling and deployment decisions. Examples and tools: image generators (Stable Diffusion, Gemini) produce stereotyped judges and legal assistants; inpainting attacks reveal memorized rare examples showing privacy risk; model choices matter — Llama 3 70B ≈ 2,290 tCO2e vs ResNet18/ImageNet ≈ 0.001 tCO2e; prefer smaller transparent models like OpenGPTx 7B (≈40% English, public data and weights), pick server locations by grid electricity factor, and apply GDPR/AI Act checks throughout the lifecycle.
AI-written summary · about
chapters (21)
- 0:00Ethics in AI development
- 1:08Q&A on AI usage
- 3:17Ethical blind spots in AI
- 4:20Biases in AI images
- 8:02Ethical risks in AI
- 10:33Fundamental rights for AI
- 12:21Implementing rights in AI
- 14:52Transparency in AI
- 17:19Bias in language models
- 20:11Privacy risks for groups
- 21:35Representation vs efficiency
- 22:41Transparent AI models
- 24:28Fairness beyond ethics
- 25:10AI carbon footprint
- 28:32Reducing AI’s impact
- 30:38Server energy efficiency
- 33:14Cost of ethical AI
- 33:36Legal compliance in AI
- 36:54Regulations driving AI
- 38:17Copyright in AI systems
- 39:41Trustworthy AI goal
related talks
01h 01mDoing Data Science with Ethics
28mKeynote, Why and how to care about ethics in ML
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
47mEthics in AI Panel
55mKeynote - Big Data's Ethical Drought
50mEthical Machine Learning
06mThe Dark Side of AI - Environmental & Social Costs Developers Must Know
32mTowards a Practical Ethics of Generative AI