Ethical ML: Who's Afraid of the Black Box Models?
summary
Argues that counterfactual fairness lets black‑box models meet GDPR and ALTAI without full mechanistic transparency, and demonstrates a working CI/CD pipeline that rejects models until a suite of counterfactual tests pass. Describes a live project on children’s welfare that records consent and contributor pledges in a blockchain ledger, exposes a public endpoint for journalists to generate counterfactual cases, and reports the approach already satisfies audits and prevented risky models from reaching production.
AI-written summary · about
chapters (9)
- 0:00Intro
- 1:26Dragons & unicorns
- 3:42Beyond model transparency & explainability
- 4:30Why should you care?
- 7:16We (data scientists) have messed things up
- 14:48We (data scientists) are fixing it
- 21:25Demo
- 33:55Conclusions
- 38:34Outro
related talks
50mEthical Machine Learning
54mEthical Machine Learning
47mEthical Machine Learning
35mBuilding Inclusive Machine Learning Models
43mIn defence of uncertainty
26m(Un)ethical Artificial Intelligence: How to Keep the AI Fair for Everyone
47mAI Ethics, Strategic Decisioning and Game Theory
27mThe ethical algorithm