Marc Salomon

Transcript

1. About the good and the bad things of [AI] Algorithms and the European AI Act Marc Salomon, Professor of Decision Science, University of Amsterdam
2. 1 WHITE BOX MODELS (= explain what happens reasonably well) BLACK BOX MODELS (= difficult to explain) Decision tree Linear regression Neural network This talk is about “the good” and “the bad” of algorithms
3. 2 What is the fear ? That we use (AI) algorithms that make decisions - intentionally or not - with far-reaching social, health, security and / or financial consequences
4. 3
5. 4 OMG, this does not happen here, …. Dutch childcare tax benefit affair. Algorithms determine who gets on the blacklist of the tax authorities and may run into big financial problems
6. 5 Harm to our democratic system – Cambridge Analytica • Floating voters were identified using Facebook data • They were presented information (via Facebook and other channels) in favour of voting Trump
7. Harm to our wallets – digital cartels 6
8. 7 But, …., there is also another fear to take very seriously Harm due to not using AI, for instance to save lives, because they can not be developed due to some laws
9. 8 A difficult balance Forbidding “the good” (= too strict rules) Continuing with “the bad” (= too loose rules)
10. 9 Extra complication in the discussion on algorithms: different people/cultures/ethical schools have different perspectives on “good” and “bad” (“good” and “bad” are subjective)
11. MIT Moral Machine to test people's perspectives
12. MIT Moral Machine to test people's perspectives
13. 12 Different “schools” in ethics think differently
14. 13 Different “schools” in ethics think differently Utilitarian approach: The greatest benefit to the greatest group. The Fairness of Justice Approach: All people should be treated the same Many other ethical schools…. Source: https://www.scu.edu/ethics/ethics-resources/ethical-decision-making/thinking-ethically/
15. 14 How to deal with this ? Different ethical schools Different applicable laws Technical limitations (math, CS, audit)
16. 15 Mathematical limitations Explainability (= how good could we explain the outcome of the algorithm to people) Quality of outcome of the algorithm
17. Europe’s answer European AI Act 16
18. Source: EU Guidelines for Trustworthy AI, Independent High Level Expert Group 17
19. Source: EU Guidelines for Trustworthy AI, Independent High Level Expert Group 18
20. Source: EU Guidelines for Trustworthy AI, Independent High Level Expert Group 19
21. Act classifies risk of applications and domains Source: Eve Gaumond, Lawfare Blog, June 2021 20
22. 21 Hurdles with the law
23. 22 Hurdles with the European Act 1. Many concepts are very subjective. Not all rational thinkers would describe “harmful discrimination” in the same way
24. 23 Hurdles with the European Act 1. Many concepts are very subjective. Not all rational thinkers would describe “harmful discrimination” in the same way 2. Translation of the legal text into mathematics is difficult. Different mathematical concepts to describe harmful discrimination.
25. 24 Hurdles with the European Act 1. Many concepts are very subjective. Not all rational thinkers would describe “harmful discrimination” in the same way 2. Translation of the legal text into mathematics is difficult. Different mathematical concepts to describe harmful discrimination. 3. Monitoring / auditing concepts, including XAI, still not in mature stage.
26. 25 Hurdles with the European Act 1. Many concepts are very subjective. Not all rational thinkers would describe “harmful discrimination” in the same way 2. Translation of the legal text into mathematics is difficult. Different mathematical concepts to describe harmful discrimination. 3. Monitoring / auditing concepts, including XAI, still not in mature stage. 4. Algorithms have very positive contributions: how to keep them ?
27. 26
28. 27
29. 28
30. 29
31. 30 Time to share some “best practices”
32. 31 Best practices: do a communication campaign
33. 32 Best practices: provide reliable information
34. 33
35. 34
36. 35 Best practices: be careful at buying algorithms
37. 36 What are your thoughts on trustworthy AI ? Marc Salomon, Professor of Decision Science Program Director MBA Big Data & Business Analytics University of Amsterdam m.salomon@uva.nl www.linkedin.com/in/marcsalomon