The Explainability Problem: Towards Understanding AI
summary
Argues that explainability remains unsolved and gives concrete failures: off‑the‑shelf saliency maps that radiologists cannot interpret, post‑hoc explainers like LIME and SHAP that disagree on feature importance (example COMPAS) and a user study where 86% of users either don’t know what to do or pick arbitrarily, adversarial examples that flip classifiers (panda → gibbon, altered stop sign), and legal pressure from the EU AI Act Article 13 pushing for measurable, user‑empowering transparency.
AI-written summary · about
chapters (17)
- 0:00Explainability problem
- 1:14Why explainability matters
- 3:00Accuracy versus explainability
- 5:46Medical AI use cases
- 6:29Adversarial vulnerabilities
- 8:13Post-hoc explanation methods
- 9:16Open problems in explainable AI
- 10:00Saliency map gaps
- 11:44Evaluating explainability
- 13:48Disagreement among explainers
- 18:02EU AI Act regulations
- 21:03Black box vs interpretable models
- 22:06Data validation vs local explanations
- 23:54Trust and future of explainable AI
- 26:00Conclusion and call to action
- 27:25Q&A on medical reasoning
- 30:17Q&A on AI legislation and trust
related talks
35mWhy Do You Need an Explainable Ai?
45mAI-404: Explanation Not Found
24mWhy and How to Build Explainability into your ML Workflow
45mAI-404: explanation not found
24mOpen and Understandable: Designing Explainable AI for Everyone
47mDecoding the algorithm - why explainability and transparency matter when buildin...
32mExplainable AI Explained for Developers
21mWhy you should care about AI interpretability