Responsible AI Challenges and Recommendations
summary
Survey of five concrete problems: COMPAS racial bias and bail prediction models that claim a 25% crime reduction / 42% prison reduction but amplify minority rates; biometric and physiognomy claims including a 70% political‑orientation classifier and patents that infer identity from faces; platform optimization that quietly underallocates gigs leading to fines (Bologna delivery case); adversarial failures and misuse of signals for trading; and huge carbon and cost footprints of LLMs at $1M–$3M per training. Recommends audit logs, systems designed for GDPR Article 22 right to contest, registered models, human control (not just human in the loop), and targeted regulation.
AI-written summary · about
chapters (8)
- 0:00Introduction to AI ethics
- 1:00Main ethical issues in AI
- 3:07Discrimination and algorithmic bias
- 11:29Facial recognition and physionomy risks
- 16:23Lack of semantic understanding
- 20:36Waste of resources and environmental impact
- 23:44Recommendations for responsible AI
- 28:15Q&A and closing remarks
related talks
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
54mEthical Machine Learning
38mSocial Implications of Bias in Machine Learning
49mAn AI with an Agenda
47mEthics & AI: Identifying & preventing bias in predictive models
21mFairness and bias in AI
01h 09mGOVERNANCE AND REGULATION OF AI FROM THE PERSPECTIVE OF AUTONOMY AND PRIVACY
47mEthical Machine Learning