Ethics and Bias in data science
summary
Claims that biased data breaks projects: Gartner predicts about 85% of AI projects underperform because of biased data, and automated voice cloning enables fraud that steals roughly $243,000 in a reported case. Examples: face recognition shows large accuracy gaps for Black women (Amazon example), machine translation reaches human parity for English–Chinese at Microsoft, and the Tay chatbot experiment collapses when exposed to toxic inputs. Mitigations cited: InterpretML and fairness tooling in the ML stack plus partner pledges and harms‑model frameworks.
AI-written summary · about
chapters (7)
- 0:00Start
- 14:31Frame Reference Problem · Al can not use common sense
- 17:24Feature engineering problem A machine does not understand how to get the right features to understand something
- 24:22How do we identify sensitive uses for Al Al systems that make automated decisions/recommendation that impact on services a person receives
- 26:04Legal Obligations - Biometrics Only Biometrics covered by - Ethically what is UK law are DNA and acceptable? fingerprints
- 34:29Bias in Al - where does it come from? Unrepresentative models of the world - Missing or wrong training data Accurate models of the world Societal biases embedded in training
- 49:44Partner Pledge
related talks
26mData Bias: The Danger Lies in the Details
01h 01mDoing Data Science with Ethics
26mHow to develop useful and unbiased AI
31mFlagging *and* Fixing Bias in ML
55mKeynote - Big Data's Ethical Drought
43mThe Elephant in your Dataset: Addressing Bias in Machine Learning
31mPanel Discussion : Ethics in Data Science
31mData Discrimination