We are the Guardians of our Future
summary
Researchers train a model on the California gang database that includes babies and 25% false entries and it still flags people as gang members; another paper claims 90% accuracy on criminality detection from 2,000 faces but uses corporate headshots for non-criminals; image benchmarks like Labeled Faces in the Wild skew white and male, causing face recognition and translation tools to misgender or erase minorities. The talk names datasets, error rates, proxy choices, and real-world misuse and asks how to test, document, and mitigate.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 0:32Gang affiliation models
- 2:51Ethical responsibility
- 4:41Data privacy issues
- 6:19AI sexual orientation
- 8:55Dangers of classification
- 10:08Criminality detectors
- 13:45Bias and automation
- 15:40Data representation problems
- 18:36Facial recognition bias
- 22:15Bias in natural language
- 25:45Feedback loops
- 28:35Technical bias mitigation
- 32:43Algorithmic transparency
- 34:21Ethics in design
- 37:02Data set limitations
- 40:13Model evaluation
- 43:50Societal surveillance
- 46:49Diverse teams and ethics
- 49:20Conclusion and call to action
related talks
01h 01mWe are the Guardians of our Future
34mDespicable machines: how computers can be assholes
01h 05mFacial recognition, a 'convenient' and 'efficient' solution looking for a problem?
05mHow to Erase Unconscious Bias From Your AI Datasets
47mEthics & AI: Identifying & preventing bias in predictive models
31mAvoiding Digital Bias
54mFast and Loose with Facial Recognition Technology: Modern Tech, Race, and Civil Liberties
31mDoes This Algorithm Have a Soul?