Lecture 9: AI Ethics (Full Stack Deep Learning - Spring 2021)
1 like9,332 views
This document discusses ethics in artificial intelligence and machine learning. It begins with an outline of the topics to be covered, including what ethics is, long-term and near-term ethical problems in AI, best practices, and where to learn more. It then covers the definitions of different ethical theories, uses "trolley problems" to illustrate ethical intuitions, and notes there is no clear consensus among ethical theories. The document discusses long-term problems like autonomous weapons, AI replacing human labor, and ensuring AI is aligned with human values and goals. Near-term problems discussed include bias in hiring models, fairness in models like COMPAS, and issues of representation in data and AI systems.
More Related Content
What's hot (13)
Similar to Lecture 9: AI Ethics (Full Stack Deep Learning - Spring 2021) (20)
More from Sergey Karayev (9)
Recently uploaded (20)
Lecture 9: AI Ethics (Full Stack Deep Learning - Spring 2021)
- 1. Full Stack Deep Learning - UC Berkeley Spring 2021 - Sergey Karayev, Josh Tobin, Pieter Abbeel Ethics
- 2. Full Stack Deep Learning - UC Berkeley Spring 2021 Preamble • This is a huge subject, spanning many disciplines and addressing many real different problems. • We, ML practitioners, need to have a student mindset, and not assume we have the answers. These are not easy problems. • I am not an expert. Excellent resources recommended at the end. 2
- 3. Full Stack Deep Learning - UC Berkeley Spring 2021 Tech Ethics Curricula • 115 university tech ethics courses analyzed • Huge variation in materials • More consensus in outcomes: ability to critique, spot issues, and make and communicate arguments. 3 SIGCSE 2020
- 4. Full Stack Deep Learning - UC Berkeley Spring 2021 One last preamble 4 https://www.newyorker.com/books/page-turner/this-is-water
- 5. Full Stack Deep Learning - UC Berkeley Spring 2021 Outline • What is ethics • Long-term ethical problems in AI • Near-term ethical problems in AI • Best practices • Where to learn more 5
- 6. Full Stack Deep Learning - UC Berkeley Spring 2021 What is ethics 6
- 7. Full Stack Deep Learning - UC Berkeley Spring 2021 Ethics? 7 Ethics ≠ Feelings Ethics ≠ Laws Ethics ≠ Societal Beliefs https://www.scu.edu/ethics/ethics-resources/ethical-decision-making/what-is-ethics/
- 8. Full Stack Deep Learning - UC Berkeley Spring 2021 Ethical Theories 8 https://kevinbinz.com/2017/04/13/ethical-theory-intro/ Divine Command - Moral behaviors are those commanded by the divine - Criticism: not much philosophy can say Virtue Ethics - Moral behaviors uphold the person's virtues - Criticism: increasing evidence that character traits are illusory Deontology (Duty) - Moral behaviors are those that satisfy the categorical imperative (e.g. don't lie, don't kill) - Criticism: unacceptable inflexibility Utilitarianism - Moral behaviors are those that bring the most good to the most people - Criticism: How to measure utility?
- 9. Full Stack Deep Learning - UC Berkeley Spring 2021 Is one a clear winner? • Professional philosophers are just about evenly split between these 9 https://philpapers.org/surveys/results.pl
- 10. Full Stack Deep Learning - UC Berkeley Spring 2021 Intuition through "Trolley Problems" 10
- 11. Full Stack Deep Learning - UC Berkeley Spring 2021 Trolley Problems 11
- 12. Full Stack Deep Learning - UC Berkeley Spring 2021 Ethical Theories 12 https://www.researchgate.net/figure/Summary-of-major-ethical-theories_tbl2_229658531
- 13. Full Stack Deep Learning - UC Berkeley Spring 2021 Ethics of Technology • Ethics change with technological progress • e.g. Industrial revolution • e.g. Right to Internet access • e.g. Birth control, surrogate pregnancy, embryo selection, artificial womb • e.g. Lab-grown meat 13
- 14. Full Stack Deep Learning - UC Berkeley Spring 2021 Questions? 14
- 15. Full Stack Deep Learning - UC Berkeley Spring 2021 Long-term problems 15
- 16. Full Stack Deep Learning - UC Berkeley Spring 2021 Autonomous Weapons • Tempting to dismiss as far-fetched at this time • But "the future is already here, just not evenly distributed" 16
- 17. Full Stack Deep Learning - UC Berkeley Spring 2021 Replacing human labor • AI and robots replacing humans in existing jobs • Good and bad! • Interesting spin: AI controlling human labor 17 Sfdpnnfoefe! tipsu!tupsz
- 18. Full Stack Deep Learning - UC Berkeley Spring 2021 Replacing humans entirely 18
- 19. Full Stack Deep Learning - UC Berkeley Spring 2021 What's common in all these long-term problems? 19
- 20. Full Stack Deep Learning - UC Berkeley Spring 2021 Alignment • "Paperclip maximizer": AGI given the goal of producing paperclips eventually turns every atom in space into paperclips • Old lesson: establishing and communicating our goals and values is hard, and technology amplifies the difficulty 20
- 21. Full Stack Deep Learning - UC Berkeley Spring 2021 Guiding principle: AI systems we build need to be aligned with our goals and values 21
- 22. Full Stack Deep Learning - UC Berkeley Spring 2021 Questions? 22
- 23. Full Stack Deep Learning - UC Berkeley Spring 2021 Deep topic • Active area of research (active at Berkeley) • Useful lens for near-term problems, too 23
- 24. Full Stack Deep Learning - UC Berkeley Spring 2021 Near-term problems 24
- 25. Full Stack Deep Learning - UC Berkeley Spring 2021 Hiring 25
- 26. Full Stack Deep Learning - UC Berkeley Spring 2021 26 https://www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias- against-women-idUSKCN1MK08G
- 27. Full Stack Deep Learning - UC Berkeley Spring 2021 Hiring • ML model to predict hiring decision given a resume 27 ML Model
- 28. Full Stack Deep Learning - UC Berkeley Spring 2021 Hiring • ML model to predict hiring decision given a resume 28 ML Model Data Ijsjoh!efdjtjpot@! Ps!kpc!qfsgpsnbodf@
- 29. Full Stack Deep Learning - UC Berkeley Spring 2021 Hiring • ML model to predict hiring decision given a resume 29 ML Model Data World Lopxo!up!cf!cjbtfe!jo!nboz!xbzt;! .!ijsjoh!qjqfmjof! .!ijsjoh!efdjtjpot! .!qfsgpsnbodf!sbujoht Tp!op!nbuufs!xibu!xf!qjdl-! pvs!ebub!jt!bmtp!cjbtfe Uifsfgpsf!pvs! npefm!jt!cjbtfe
- 30. Full Stack Deep Learning - UC Berkeley Spring 2021 Hiring • ML model to predict hiring decision given a resume 30 Action Tpvsdjoh@!Epvcmf.difdljoh!ivnbo! efdjtjpo@!Ijsjoh@ ML Model Data World
- 31. Full Stack Deep Learning - UC Berkeley Spring 2021 Hiring • ML model to predict hiring decision given a resume 31 Action Jo!nboz!dbtft-!bnqmjgzjoh! fyjtujoh!cjbtft" ML Model Data World
- 32. Full Stack Deep Learning - UC Berkeley Spring 2021 Hiring • ML model to predict hiring decision given a resume • Amplifying existing biases is not aligned with our goals and values! 32 Action Jo!nboz!dbtft-!bnqmjgzjoh! fyjtujoh!cjbtft" ML Model Data World
- 33. Full Stack Deep Learning - UC Berkeley Spring 2021 Questions? 33
- 34. Full Stack Deep Learning - UC Berkeley Spring 2021 Fairness 34
- 35. Full Stack Deep Learning - UC Berkeley Spring 2021 COMPAS • Goal: predict recidivism, such that judges can consult 1-10 score in pre-trial sentencing decisions. • Motivation: be less biased than humans • Solution: • Gather data • Exclude protected class attributes (race, etc) • Ensure that our model's score corresponds to same probability of recidivism across all groups 35 Correctional Offender Management Profiling for Alternative Sanctions
- 36. Full Stack Deep Learning - UC Berkeley Spring 2021 And yet! 36 https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
- 37. Full Stack Deep Learning - UC Berkeley Spring 2021 21 Fairness Definitions 37 Aravind Narayanan - 21 Fairness Definitions 🙏
- 38. Full Stack Deep Learning - UC Berkeley Spring 2021 Statistical Bias • Statistical bias: Difference between estimator's expected value and the true value • In this sense, COMPAS scores are not biased, w.r.t re-arrest • But is it an adequate fairness criterion? Is it aligned with our values? 38 https://www.washingtonpost.com/news/monkey-cage/wp/2016/10/17/can-an-algorithm-be-racist-our-analysis-is-more-cautious- than-propublicas/ Jnqpsubou!dbwfbu"!Xf!pomz! ibwf!ebub!gps!bssftut-!opu! dsjnft!dpnnjuufe/!Uifsf!nbz! xfmm!cf!cjbt!jo!bssftut/
- 39. Full Stack Deep Learning - UC Berkeley Spring 2021 Different perspectives on fairness • What do different stakeholders want from the classifier? • Decision-maker: of those I've labeled high risk, how many recidivated? • Predictive value: TP / (TP + FP) • Defendant: what's the probability I'll be incorrectly classified as high risk? • False positive rate: FP / (FP + TN) • Society: is the selected set demographically balanced? • Demographic parity 39 Aravind Narayanan - 21 Fairness Definitions
- 40. Full Stack Deep Learning - UC Berkeley Spring 2021 Group Fairness • Do outcomes differ between groups (e.g. demographic), which we have no reason to believe are actually different? • Motivation of the Pro-Publica article 40 Aravind Narayanan - 21 Fairness Definitions Gbmtf! Qptjujwf Gbmtf! Ofhbujwf
- 41. Full Stack Deep Learning - UC Berkeley Spring 2021 No "correct" group fairness definition 41 Aravind Narayanan - 21 Fairness Definitions
- 42. Full Stack Deep Learning - UC Berkeley Spring 2021 It gets worse... • Okay, then let's just pick most important two metrics (FPR and FNR), and allow the model to use protected class attributes. • Now we fail individual fairness -> • Fine, let's just pick one! (e.g. equal FPR) • We still sacrifice some utility (e.g. public safety, or number of defendants released) 42 Aravind Narayanan - 21 Fairness Definitions
- 43. Full Stack Deep Learning - UC Berkeley Spring 2021 Interactive Example 43 https://research.google.com/bigpicture/attacking-discrimination-in-ml/
- 44. Full Stack Deep Learning - UC Berkeley Spring 2021 By the way... 44 Aravind Narayanan - 21 Fairness Definitions
- 45. Full Stack Deep Learning - UC Berkeley Spring 2021 Trade-offs 45 Aravind Narayanan - 21 Fairness Definitions
- 46. Full Stack Deep Learning - UC Berkeley Spring 2021 Seeing the water 46
- 47. Full Stack Deep Learning - UC Berkeley Spring 2021 47 https://www.mobilizegreen.org/blog/2018/9/30/environmental-equity-vs-environmental-justice-whats-the-difference
- 48. Full Stack Deep Learning - UC Berkeley Spring 2021 48 https://www.mobilizegreen.org/blog/2018/9/30/environmental-equity-vs-environmental-justice-whats-the-difference
- 49. Full Stack Deep Learning - UC Berkeley Spring 2021 49 https://www.mobilizegreen.org/blog/2018/9/30/environmental-equity-vs-environmental-justice-whats-the-difference
- 50. Full Stack Deep Learning - UC Berkeley Spring 2021 Questions? 50
- 51. Full Stack Deep Learning - UC Berkeley Spring 2021 Representation 51
- 52. Full Stack Deep Learning - UC Berkeley Spring 2021 Representation 52 https://twitter.com/nke_ise/status/897756900753891328
- 53. Full Stack Deep Learning - UC Berkeley Spring 2021 Not a new problem, sadly 53
- 54. Full Stack Deep Learning - UC Berkeley Spring 2021 ...or a problem in just our field 54 Recent improvement!
- 55. Full Stack Deep Learning - UC Berkeley Spring 2021 Large part of the solution 55 https://www.nytimes.com/2021/03/15/technology/artificial-intelligence-google-bias.html
- 56. Full Stack Deep Learning - UC Berkeley Spring 2021 Another example 56 Recent improvement!
- 57. Full Stack Deep Learning - UC Berkeley Spring 2021 Translation 57 Aravind Narayanan - 21 Fairness Definitions Recent improvement!
- 58. Full Stack Deep Learning - UC Berkeley Spring 2021 Word Embeddings • Word2Vec introduced vector math on word embeddings • Reveal harmful biases encoded in our language corpora • Potential solution: de-bias at training time, but at least make user aware 58
- 59. Full Stack Deep Learning - UC Berkeley Spring 2021 GPT-3 59
- 60. Full Stack Deep Learning - UC Berkeley Spring 2021 Recent Language Models • Paper apparently led to authors being fired by Google 60 http://faculty.washington.edu/ebender/papers/Stochastic_Parrots.pdf
- 61. Full Stack Deep Learning - UC Berkeley Spring 2021 Seeing the water • Whether language models should reflect the world as it is in the data, or as society believes it should be, depends on what use they are applied to. • And when we do want them to reflect the world as it should be, do we agree on what that is? 61
- 62. Full Stack Deep Learning - UC Berkeley Spring 2021 Face Recognition • Old context: no expectation of privacy in public • New context: "in public" now means "on any street in any city, or on any website on the internet" • Is it ethical to work on this? • Is it a problem if it does not work as well on some ethnicities? 62 https://fairmlbook.org/introduction.html
- 63. Full Stack Deep Learning - UC Berkeley Spring 2021 Questions? 63
- 64. Full Stack Deep Learning - UC Berkeley Spring 2021 Best practices 64
- 65. Full Stack Deep Learning - UC Berkeley Spring 2021 What do practitioners need? 65 • support in fairness-aware data collection and curation • overcoming teams’ blind spots • implementing more proactive fairness auditing processes • auditing complex ML systems • deciding how to address particular instances of unfairness • addressing biases in the humans embedded throughout the ML development pipeline 2019
- 66. Full Stack Deep Learning - UC Berkeley Spring 2021 Some suggestions • Ethical risk sweeping treat like cybersecurity penetration testing • Expanding the ethical circle whose interests, desires, experiences, values have we just assumed instead of consulted? • Think about the terrible people Who might abuse, steal, weaponize what we build? What incentives are we creating? • Closing the loop Remember that this is not a process to complete and forget. Set up ways to keep improving. 66 https://www.youtube.com/watch?v=av7utkFXbU4
- 67. Full Stack Deep Learning - UC Berkeley Spring 2021 Model Cards 67 https://modelcards.withgoogle.com/face-detection
- 68. Full Stack Deep Learning - UC Berkeley Spring 2021 Bias Audit 68 http://www.datasciencepublicpolicy.org/projects/aequitas/
- 69. Full Stack Deep Learning - UC Berkeley Spring 2021 69 http://www.datasciencepublicpolicy.org/projects/aequitas/
- 70. Full Stack Deep Learning - UC Berkeley Spring 2021 70 AI bias is rooted in societal inequities, manifested in data, and amplified by AI. Unbiased societal AI is an impossible goal for a single project to achieve, but greatly mitigating its harmful effects is not. Visualization courtesy of Eric Wang
- 71. Full Stack Deep Learning - UC Berkeley Spring 2021 71 Build a diverse AI team to bring lived experience to their work Display insights about bias contextually and intuitively in product Make AI fairness and transparency a key component of our work Visualization courtesy of Eric Wang
- 72. Full Stack Deep Learning - UC Berkeley Spring 2021 A professional code of ethics? 72
- 73. Full Stack Deep Learning - UC Berkeley Spring 2021 Medicine 73
- 74. Full Stack Deep Learning - UC Berkeley Spring 2021 Armed Forces 74
- 75. Full Stack Deep Learning - UC Berkeley Spring 2021 75
- 76. Full Stack Deep Learning - UC Berkeley Spring 2021 Questions? 76
- 77. Full Stack Deep Learning - UC Berkeley Spring 2021 Learn more 77
- 78. Full Stack Deep Learning - UC Berkeley Spring 2021 Experts • Practical Data Ethics course from Rachel Thomas: https://ethics.fast.ai • Single lecture from Fast.ai 2020 course: https:// www.youtube.com/watch?v=krIVOb23EH8 78
- 79. Full Stack Deep Learning - UC Berkeley Spring 2021 79 https://fairmlbook.org
- 80. Full Stack Deep Learning - UC Berkeley Spring 2021 80 https://dssg.github.io/fairness_tutorial/
- 81. Full Stack Deep Learning - UC Berkeley Spring 2021 81
- 82. Full Stack Deep Learning - UC Berkeley Spring 2021 82
- 83. Full Stack Deep Learning - UC Berkeley Spring 2021 Thank you! 83