SlideShare a Scribd company logo

A Tutorial to AI Ethics - Fairness, Bias & Perception

Dr. Kim (Kyllesbech Larsen), profile picture
Dr. Kim (Kyllesbech Larsen)

The document discusses ethical concerns related to artificial intelligence (AI), focusing on public sentiment and potential job impacts, as indicated by various surveys. It presents ethical dilemmas faced by AI designers, particularly in autonomous vehicle programming decisions, and highlights biases in AI, particularly in criminal sentencing algorithms. It emphasizes the necessity for ethical frameworks in AI development that consider cultural and societal norms to address issues of bias and fairness.

Technology
Related topics:
AI Ethics Overview • Facial Recognition
1 of 56
1
2
Most read
3
4
5
6
7
Most read
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
Most read
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
Ethics for Artificial Intelligent Beings. AI Ethics Workshop, 25th of June, 2018, Bonn, Germany. Dr. Kim Kyllesbech Larsen, Deutsche Telekom. Do we need to care about it? AI
2 Dr. Kim K. Larsen / How do we Humans feel about AI? What about Us?
HOW DO YOU FEEL ABOUT A.I.? 16% 36% 48% Negative Neutral Positive 20% of respondents are enthusiastic about AI. 15% are uncomfortable or scared about AI. Millennials are significantly more negative towards AI. SurveyMonkey “Artificial Intelligence & Human Decision Making Sentiment Survey “ (November 2017); 467 responses. (Note: this data does not include Millennial statistics)
DO YOU BELIEVE THAT AI COULD IMPACT YOU OR YOUR FRIENDS JOBS? Yes, 47% No, 53% Yes, 20% No, 80% Do you believe your job could be replaced by an AI? Thinking of your friends, do you believe their jobs could be replaced by an AI? SurveyMonkey “Millennials – Digital Challenges & Human Answers Survey “ (March 2018), average age 28.8 years.
How do you think AI’s will impact your child’s or children’s future in terms of job & income? SurveyMonkey “Artificial Intelligence & Human Decision Making Sentiment Survey “ (November 2017); 467 responses. 20% 16% 64% Worse Same as Today Better 31% 23% 46% Worse Same as Today Better Men with Children above 18 yrs of Age Men with Children under 18 yrs of Age
Note: The gender bias on this slide is completely intended.
A Tutorial to AI Ethics - Fairness, Bias & Perception
A DIE LIVE • AN AI-BASED AUTONOMOUS CAR SLIPS ON AN ICY MOUNTAIN ROAD OVER TO OPPOSITE SIDE OF ROAD WHERE A NORMAL CAR IS APPROACHING. • THE AI IS PROGRAMMED TO NOT DELIBERATELY CAUSE INNOCENT BYSTANDERS HARM. • THE AI CONTINUES THE CARS DIRECTION 200 METER DOWN THE ROCKY VALLEY. A FAMILY OF 4 PARISH. • “SAVING” THE 1 PASSENGER OF THE APPROACHING CAR.
B DIE LIVE • AN AI-BASED AUTONOMOUS CAR SLIPS ON AN ICY MOUNTAIN ROAD OVER TO OPPOSITE SIDE OF ROAD WHERE A NORMAL CAR IS APPROACHING. • THE AI IS PROGRAMMED TO SAVE ITS PASSENGERS AT ALL COST FROM MORTAL DANGER IF THE LIKELIHOOD OF SUCCESS IS HIGHER THAN 50%. • THE AI COLIDES WITH THE APPROACHING CAR & SAVES ITS 4 PASSENGERS. • THE APPROACHING NORMAL CAR IS PUSHED 200 METERS DOWN INTO THE ROCKEY VALLEY & ITS 1 PASSENGER PERISH.
YOU ARE THE AI DESIGNER! Q1: SHOULD YOU PROGRAM THE AUTONOMOUS CAR AI TO SAVE ITS DRIVER (& POSSIBLE PASSENGERS) AT ALL “COST”? Q2: WOULD YOU DEFINE A “COST” THRESHOLD?
Immanuel Kant Jeremy Bentham  Deontological Ethics  Duty-based.  Rule-based (e.g., Rule of Law).  Asimov’s Laws.  10 Commandments.  Golden Rules.  Moral imperative.  Utilitarian Ethics  Consequentialism.  The “greater” good.  "it is the greatest happiness of the greatest number that is the measure of right and wrong.”
An ethical framework will depend on the cultural, political & religious background of the issuer.
AI = ? = ? = ? = ? = ?AI AI AI AI AI = ? AI For an AI Ethical Framework Design Does religious background matter? Does socio-cultural background matter?
Societal Ethics (Public, Policy & Law Makers) Business Ethics (Corporations) Data Universe (all data cumulated) Part of the Data Universe deemed permissible by societal norms for algorithmic processing (analysis). Note that all of what might be deemed permissible from a societal perspective might not be acceptable from a business/commercial ethical perspective, Part of the Data Universe deemed permissible by business ethical norms for algorithmic processing (analysis). Note this can (should) never be larger than what is in general acceptable to society or the rule of law. Algorithm sub-space … that could be acting on the Data Universe under given societal or business ethical guidelines Part of the Data Universe that has been deemed impermissible to process algorithmically by Society incl. Businesses. E.g, due to privacy concern, un-ethical (within societal norms & laws) use, etc..
What Ethics do you Wish for?
A Tutorial to AI Ethics - Fairness, Bias & Perception
KIM = WOMAN NURSE = WOMAN DOCTOR = MAN COOKING = WOMAN MANAGER = MAN HIGH RISK LOW RISK https://www.propublica.org/article/machine- bias-risk-assessments-in-criminal-sentencing “Men also like shopping” by Jieyu Zhao et al. By Joy Buolamwini et al https://newrepublic.com/article/144644/turns- algorithms-racist As AIbecomes more andmorecomplex, it can become difficult for even its own designers understand why it acts the way it does.
Cognitive biases Statistical bias Contextual biases Bias is a disproportionate weight in favor of or against one thing, person, or group compared with another, usually in a way considered to be unfair. ℬO MO = E MO − O O: Observation. M: Model estimator of observation O. E: Expected value (long-run average) B: Bias of model of O relative to O. Hundreds of cognitive biases, e.g., Anchoring, Aavailability, Confirmation bias, Belief bias, framing effect, etc… (See: https://en.wikipedia.org/wiki/List_of_cognitive_biases) Academic bias, Experimenter bias, Educational bias, Religious bias, cultural bias, etc… (See: https://en.wikipedia.org/wiki/Bias#Contextual_biases)
Basic on fairness & statistical bias ℬ = P Y = y X = x, G = b) − P Y = y X = x, G = w) X = x ∈ ℝn varables of direct interest i. e. , covariates for a given score S = S(x). Group G ∈ b, w of which individuals belong, e.g., b = black (African-Americans), w = white (Caucasians). Binary outcome indicator Y = y ∈ 0,1 depending on S, e.g., Y = 1 indicates individual High Risk (S ≥ 4) of re-offend. Y = 0 indicates individual Low Risk (S < 4) of re-offend. ℬ = 0 then score S is well calibrated unbiased . ℬ ≠ 0 then score S is said to be biased. Based on Chouldechova (2017) Fair prediction with disparate impact A study of bias in recidivism prediction instruments.
Bias case study (1 of 4) Machine Bias in Risk Assessments of Criminal Sentencing US Distribution Arrests Likelihood of being arrested Population Total Total Male Population reference Caucasians 61% 57% 2.4% 3.7% 1.5% African-Americans 13% 27% 5.2% 7.9% 0.7% Others 26% 16% 1.6% 2.4% 0.4% Total 100% 100% 2.6% 3.9% 2.6% Race Arrest Likelihood within Race • In USA there are ca. 4.5× more Caucasians than African-Americans. • Still African-Americans are more than 2+ times more likely to be arrested. • Ca. 11% more African-Americans males (79%) where repeat offender compared to Caucasians (68%). • Ca. 3% more African-Americans males (18%) where violent recidivist compared to Caucasians (15%). Propublica found : a. Odds of a African-American male being assessed to have a high risk of recidivism is double that for a Caucasian male. b. Substantially higher amount (2× - 3×) of False Positives among AAs compared to Cs.
Bias case study (2 of 4) Machine Bias in Risk Assessments of Criminal Sentencing High Risk of Recidivism differences between African-Americans & Caucasians 𝓑 The difference in getting a High Risk Score if you are a Black Male vs White Male. P(High Risk | Male, African-American) – P(High Risk | Male, Caucasian) > 0
Bias case study (2 of 4) Machine Bias in Risk Assessments of Criminal Sentencing High Risk of Recidivism differences between African-Americans & Caucasians 𝓑 The difference in getting a High Risk Score if you are a Black Male vs White Male and being less than 25 yrs old P(High Risk | Male, Age<25yrs, African-American) – P(High Risk | Male, Age<25yrs, Caucasian) > 0
Bias case study (2 of 4) Machine Bias in Risk Assessments of Criminal Sentencing No statistically significant difference High Risk of Recidivism differences between African-Americans & Caucasians ℬ = P Y = 1 𝐗 = 𝐱, G = African − American) − P Y = 1 𝐗 = 𝐱, G = Caucasian) 𝑌 = 1 𝑓𝑜𝑟 4 ≤ 𝑆(𝑥) ≤ 10 𝐻𝑖𝑔ℎ 𝑅𝑖𝑠𝑘 𝑜𝑓 𝑅𝑒𝑐𝑖𝑑𝑖𝑣𝑖𝑠𝑚 0 𝑓𝑜𝑟 𝑆 𝑥 < 4 𝐿𝑜𝑤 𝑅𝑖𝑠𝑘 𝑜𝑓 𝑅𝑒𝑐𝑖𝑑𝑖𝑣𝑖𝑠𝑚 X = x ∈ 𝑀𝑎𝑙𝑒, 𝐿𝑒𝑠𝑠 − 𝑡ℎ𝑎𝑛 − 25, 𝐹𝑒𝑙𝑜𝑛𝑦, 𝑃𝑟𝑖𝑜𝑟𝑠, 𝐽𝑢𝑣𝑒𝑛𝑖𝑙𝑒 𝑃𝑟𝑖𝑜𝑟𝑠, 𝑉𝑖𝑜𝑙𝑒𝑛𝑡 𝑅𝑒𝑐𝑖𝑑 𝓑
44 11 15 31 25 14 14 47 62 8 15 15 27 21 7 45 69 5 17 9 2 26 2 70 33 11 17 39 African-American Caucasian Males Below 25 yrs Below 25 yrs Above 25 yrs Above 25 yrs 0 23 0 77 6 33 7 54 16 14 9 69 51 6 26 17 2 26 2 70 33 11 17 39 2 26 2 70 33 11 17 39 2 Yr recid. No 2 Yr recid. 2 Yr recid. No 2 Yr recid. 2 Yr recid. No 2 Yr recid. 2 Yr recid. No 2 Yr recid. 0 14 0 86 0 29 0 71 0 13 3 84 0 0 33 67 12 42 15 30 78 0 22 0 75 0 25 0 93 0 7 0 No Juv. Priors + No Priors Juvenile Priors Juvenile Priors Juvenile Priors Juvenile Priors No Juv. Priors + No Priors No Juv. Priors + No Priors No Juv. Priors + No Priors Bias case study (3 of 4) TN LOW RISK FP HIGH RISK FN LOW RISK TP High RISK Note: the binary classification model used here does include Race to predict recidivism risk. Ignoring race in the model does not lead to substantially less biased results. FP → A Human is incorrectly (“falsely”) assessed to have a High Risk of re-offending → Gets a substantially stricter treatment. FN → A Human is incorrectly (“falsely”) assessed to have a Low Risk of re-offending → Gets a substantially lighter treatment. Might result in serious crime. Confusion Matrix (all normalized)
Bias case study (4 of 4) Debiasing strategies (non exhaustive). 25 TN LOW RISK 14 FP HIGH RISK 14 FN LOW RISK 47 TP High RISK 62 TN LOW RISK 8 FP HIGH RISK 15 FN LOW RISK 15 TP High RISK African-American Males Caucasian Males Binary Classification Model of Low vs High Risk recidivism. Model include; Race (African-American, Caucasian & Other), Gender (female / male), age category (below & above 25), Charge Degree (misdemor & felony), #Juvenile prior count (≥0). #Prior counts (≥0), 2 Yr recidivism (0/1) & Violent Recidivism (0/1). No de-biasing 28 TN LOW RISK 10 FP HIGH RISK 19 FN LOW RISK 42 TP High RISK 58 TN LOW RISK 9 FP HIGH RISK 15 FN LOW RISK 18 TP High RISK African-American Males Caucasian Males Binary Classification Model of Low vs High Risk recidivism. Model note; Race has been taken out of the model. Remove race from model 32 TN LOW RISK 8 FP HIGH RISK 24 FN LOW RISK 36 TP High RISK 57 TN LOW RISK 5 FP HIGH RISK 22 FN LOW RISK 17 TP High RISK African-American Males Caucasian Males Binary Classification Model of Low vs High Risk recidivism. Model note; Race has been taken out of the model. The training data has been rebalanced from (CC, AA, OTH)=(34%, 51%, 15%) to the demographic blend (61%, 13%, 26%). However, it does result in less training data as the proportion of African- Americans (AA) are dramatically reduced, i.e., likely hurting accuracy. Remove race from model & re-balance race mix to demographic blend FPR = 36 FNR = 22 FPR = 11 FNR = 43 FPR = 15 FNR = 32 FPR = 9 FNR = 51 FPR = 28 FNR = 30 FPR = 17 FNR = 45
The “Unfairness Law” On group fairness. If a model satisfy predictive parity, but the prevalence differs between groups, then that model cannot achieve equal False Positive & False Negative Rates across those groups. See: Chouldechova (2017) Fair prediction with disparate impact A study of bias in recidivism prediction instruments Base Rate P(Y=1 | G = g) Positive Predictive Value PPV = TP / (TP + FP) Not all fairness criteria can be satisfied at the same time!
Is Google Translate misogynist? English → Hungarian → English He is a Nurse. She Is a Medical Doctor. Note: Hungarian language does not have gender-specific pronouns and lacks grammatical gender. If starting point is “The Woman is a Doctor, the Man is a Nurse”, Google translate from and back to English via Hungarian will work better. Note the Hungarian Ő has the meaning of She/He You will get the same translation bias with “He is an Assistant. She is a Manager.”
Father is to a doctor as a Mother is to a nurse Man is to computer programmer as Woman is to homemaker Boy is to gun as Girl is to Doll Man is to Manager as Woman is to Assistant Gender biases. https://developers.google.com/machine-learning/fairness-overview/
Man is to Manager as Woman is to Assistant Approach to de-biasing biased representations. bias non-bias Assistant Manager Man Woman He She Bolukbasi et al (2016), “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings”, 30th Conferenceon Neural Information Processing Systems (NIPS2016), Barcelona, Spain DefinitionalNeutralWords Word Embeddings e ∈ ℝ 𝑛 Softmax Word Embeddings e ∈ ℝ 𝑛 3 Bn Words 3 mn 300-dim English words w2vNEWS vector • Find bias direction based on difference between definitional opposing word vectors (e.g., he-she, male-female, …) • Neutralized non-definitional words by projecting them on to the non-bias axis.
Man is to Manager as Woman is to Assistant Approach to de-biasing biased representations. bias non-bias Assistant Manager Man Woman He She Bolukbasi et al (2016), “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings”, 30th Conferenceon Neural Information Processing Systems (NIPS2016), Barcelona, Spain Woman is closer to Assistant than Man Woman is closer to Assistant than Man Girl Boy DefinitionalNeutralWords Word Embeddings w ∈ ℝ 𝑛 • Find bias direction based on difference between definitional opposing word vectors (e.g., he-she, male-female, …) • Neutralized non-definitional words by projecting them on to the non-bias axis. • Equalize pairs of definitional words to be equidistant to Neutral Words.
Man is to Manager as Woman is to Assistant Approach to de-biasing biased representations. bias non-bias Assistant Manager Man Woman He She Bolukbasi et al (2016), “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings”, 30th Conferenceon Neural Information Processing Systems (NIPS2016), Barcelona, Spain Equalize pairs to be equidistant to non-biased descriptor Girl Boy DefinitionalNeutralWords Word Embeddings w ∈ ℝ 𝑛 • Find bias direction based on difference between definitional opposing word vectors (e.g., he-she, male-female, …) • Neutralized non-definitional words by projecting them on to the non-bias axis. • Equalize pairs of definitional words to be equidistant to Neutral Words.
"Gay faces tended to be gender atypical," the researchers said. "Gay men had narrower jaws and longer noses, while lesbians had larger jaws." Wang, Y., & Kosinski, M. (in press). Deep neural networks are more accurate than humans at detecting sexual orientation from facial images. Journal of Personality and Social Psychology DEEP NEURAL NETWORKS CAN DETECT SEXUAL ORIENTATION FROM FACES
ACCURACY = 80% What’s the likelihood you are Gay given your have been “diagnosed” Gay; P(Gay │Positive Detection) = P(G) × P( PD│G) divided by ( P(G)×P( PD│G) + P(¬G )×P( PD│ ¬G ) NOTE: This estimate given here is illustrative and possible wrong. As Wang, Y., & Kosinski has not provided other numbers than their 81% accuracy and supposedly a 23% False Negative rate, i.e., algorithm predicts a Gay man to be Straight. TRUE NEGATIVE 40% TRUE POSITIVE 40% FALSE POSITIVE 10% FALSE NEGATIVE 10% ActualClass Predicted Class Males Only GAY STRAIGHT GAYSTRAIGHT P(Gay│Positive Detection) = 2% × 40% / (2% × 40% + 98% × 10%) ≈ 8% and …. after 5 positive detections the likelihood is ≈ 73% (Don’t take this analysis too serious! … I don’t). Own guestimates FP → A Human is incorrectly (“falsely”) assessed to be Gay → Can lead to severe repercussions for the individual. FN → A Human is incorrectly (“falsely”) assessed to Straight → Unlikely to have any impact. TP → A Human is correctly (“truly”) assessed to be Gay → Can lead to severe repercussions for the individual.
? 64x64x3 Female German Telekom DNN Architecture e.g., 128/64/32/1 (4 Layers) Trained on 6,992+ LinkedIn pictures. TRUE POSITIVES Male Polish Vodafone FALSE NEGATIVE What’s your Gender, Nationality & Employer. How much does your face tell about you?
What’s your Gender? MANp – MANa TRUE NEGATIVE WOMANp - WOMANa TRUE POSITIVE WOMANp - MANa FALSE POSITIVE MANp - WOMANa FALSE NEGATIVE ActualClass Predicted Class On Test Data 0.45 0.32 0.16 0.07 Predicted Class ACC = 77% PRE = 74% REC = 87%
36 The Cost of Machine Error. (and of Human error) 20 Million Muslims in EU. Expect < 1,300 active (Muslim) terrorists in EU* ~ 1 in 15 Thousand. (note: 5,000+ Europeans estimated to have travelled to Iraq & Syria by end of 2015) Assume a sophisticated model** gives the following: Thsd. FALSE (0) TRUE (1) FALSE (0) 19,978 9.0 TRUE (1) 0.3 1.0 Costly police work and time spend on the wrong people. Predicted Actual (*) 687 (suspects) were arrested for terrorism related-offences in 2015 (source: Europol TE-SAT 2016 report). (**) e.g., Bayesian machine learning models, Deep learning methodologies, social network analysis (e.g., social physics). STILL BOOM! ? ACCURACY = 99.99% ERROR = 0.01%
Performance of Learning Machines. (the cost of machine error vs human error) TRUE NEGATIVE (TN) TRUE POSITIVE (TP) FALSE POSITIVE (FP) FALSE NEGATIVE (FN) Actual Class Predicted Class Precision = TP TP + FP TP TP + FN Recall = TN TN + FN Accuracy = TP + TP + FP + F1-Score = + 2 1 1 Prec. Recl.  Could be Very Costly! Could be Very Costly!    (Ex. 10%) (Ex. 77%) (Ex. 99.99%) (Ex. 18%) Note: the structure of Python sklearn.metrics confusion_matrix.is used. (Positive Predictive Value) (Sensitivity)
Ethical AI architectures – Illustration (1 of 2). Including Bias & Fairness checks with corrections (e.g., de-biasing)
Ethical AI architectures – Illustration (2 of 2). Vanderelst, D. and Winfield, A. (2018). An architecture for ethical robots inspired by the simulation theory of cognition. Cognitive Systems Research, 48, pp.56-66. Roboticist Alan Winfield of Bristol Robotics Laboratory in the UK built an ethical trap for a robot following Asimov’s Laws. https://www.youtube.com/watch?v=jCZDyqcxwlo 42% of the time the Robot took so long to decide action that the “Human” perished.
40 Dr. Kim K. Larsen / How do we Humans feel about AI? What about Us?
41 • 4 AI-related surveys • Paid responses. • Social media responses. • Error-margin 4% - 6%.  US population focused.  18 – 75 Yrs old.  HH Income > 75k US$.  Millennials 18 – 40 yrs old (*) Median US HH income in 2016 was 58k pa. A little more than 30% of US households earn 75k or more annually (2016).  Do weHumans trust AIs?  Acceptance of Artificial Intelligence inCorporate Decision Making.  AIEthics. AI Strategy&Policy Aistrategyblog.com
THE GOOD & THE VERY BAD “How to train your ai chatbot to become evil in less than 24 hours”.
Humans are increasingly dependent on digital technologies. “Today, everything & more that makes us us are replicated in the digital world.
A Tutorial to AI Ethics - Fairness, Bias & Perception
10% 39% 28% 13% 10% Exclusive Frequently Sometime Rarely Never How often do you rely on news from social media ? ~50% 7% 14% 43% 22% 14% Very High High About half appears truthfull Low Very Low What is your trust level in what you read on social media ? ~80%
Is it acceptable that your personal data, residing on social media platforms, are used for ADVERTISING? INFLUENCING ? Yes, 18% No, 82% Yes, 33% No, 67% SurveyMonkey “Millennials – Digital Challenges & Human Answers Survey “ (March 2018), average age 28.8 years.
ARE YOU BEING INFLUENCED?
4% 52% 44% Below average Average Above average ~44% Thinking of your friends, do you believe their opinions & values are negatively influenced by social media? How would you characterize your ability to detect fake news compared to your friends? 50% 40% 10% Frequently Sometimes Rarely ~50%
4 Dr. Kim K. Larsen / How do we Humans feel about AI? A I
HOW DO YOU FEEL ABOUT A.I.? Millennials (18 – 38 yrs old). 31% 45% 24% Negative Neutral Positive ~24% 23% 41% 36% Negative Neutral Positive ~36% SurveyMonkey “Millennials – Digital Challenges & Human Answers Survey “ (March 2018), average age 28.8 years.
51 How do you feel about A.I.? (All Groups) SurveyMonkey “Artificial Intelligence & Human Decision Making Sentiment Survey “ (November 2017); 467 responses. 18% 44% 38% Negative Neutral Positive 14% 30% 57% Negative Neutral Positive 16% 44% 41% Negative Neutral Positive WOMEN WITH CHILDREN UNDER 18 YRS OF AGE 8% 27% 65% Negative Neutral Positive MEN WITH CHILDREN UNDER 18 YRS OF AGE
Would you trust an AI with a critical corporate decision? SurveyMonkey “Artificial Intelligence & Human Decision Making Sentiment Survey “ (November 2017); 467 responses. AI “Allergy” is a Real Corporate Ailment! 19% 43% 25% 12% 0% Never Infrequently About half the time Frequently Always Would you trust a critical corporate decision made by an AI?12% 2% 9% 36% 47% 6% Never Infrequently About half the time Frequently Always Would you trust a critical corporate decision made by a fellow human expert or superior?53%
53Dr. Kim K. Larsen / How do we Humans feel about AI? We hold AI-s to much stricter standards than our fellow humans. Who do you trust the most with corporate decisions … Your fellow Human or your Corporate AI? 2% 12% 33% 49% 4 % Never Infrequently About half the time Frequently Always Would you trust a corporate decision made by a fellow human whos sucess rate is better than 70%?53% 11% 32% 40% 16% 1% Never Infrequently About half the time Frequently Always Would you trust a critical corporate decision made by an AI which success rate is better than 70%?17% Think a minute about how forgiving we are with the accident made by human drivers versus self-driving cars! SurveyMonkey “Artificial Intelligence & Human Decision Making Sentiment Survey “ (November 2017); 467 responses.
Do you trust that companies using AI have your best interest in mind? ≈ 25% 75% Yes No ≈
A Tutorial to AI Ethics - Fairness, Bias & Perception
THANK YOU!Acknowledgement Many thanks to Viktoria Anna Laufer in particular and other colleagues who have contributed with valuable insights, discussions & comments throughout this work. Also I would like to thank my wife Eva Varadi for her patience during this work. Contact: Email: kim.larsen@telekom.hu Linkedin: www.linkedin.com/in/kimklarsen Blogs: www.aistrategyblog.com & www.techneconomyblog.com Twitter: @KimKLarsen

More Related Content

PDF
Ai Ethics
sparks & honey
 
PPTX
Technology for everyone - AI ethics and Bias
Marion Mulder
 
PPTX
Implementing Ethics in AI
Pekka Abrahamsson / Tampere University
 
PPT
Ethics and AI
Thodoris Panagiotopoulos
 
PDF
Ethical issues facing Artificial Intelligence
Abdelhak
 
PPTX
Ethical Considerations in the Design of Artificial Intelligence
John C. Havens
 
PDF
Introduction to the ethics of machine learning
Daniel Wilson
 
PDF
Ethics in the use of Data & AI
Kalilur Rahman
 
Ai Ethics
sparks & honey
 
Technology for everyone - AI ethics and Bias
Marion Mulder
 
Implementing Ethics in AI
Pekka Abrahamsson / Tampere University
 
Ethics and AI
Thodoris Panagiotopoulos
 
Ethical issues facing Artificial Intelligence
Abdelhak
 
Ethical Considerations in the Design of Artificial Intelligence
John C. Havens
 
Introduction to the ethics of machine learning
Daniel Wilson
 
Ethics in the use of Data & AI
Kalilur Rahman
 

What's hot (20)

PPTX
Responsible AI in Industry (ICML 2021 Tutorial)
Krishnaram Kenthapadi
 
PPTX
Generative AI Risks & Concerns
Ajitesh Kumar
 
PPTX
AI and the Impact on Cybersecurity
Graham Mann
 
PPTX
Responsible AI
Neo4j
 
PDF
How do we train AI to be Ethical and Unbiased?
Mark Borg
 
PDF
Exploring Opportunities in the Generative AI Value Chain.pdf
Dung Hoang
 
PDF
Algorithmic Bias: Challenges and Opportunities for AI in Healthcare
Gregory Nelson
 
PDF
What is Artificial Intelligence | Artificial Intelligence Tutorial For Beginn...
Edureka!
 
PPT
Artificial intelligence
funpathshala
 
PDF
Responsible AI
Neo4j
 
PDF
9 Examples of Artificial Intelligence in Use Today
IQVIS
 
PDF
An Introduction to Generative AI - May 18, 2023
CoriFaklaris1
 
PDF
GENERATIVE AI, THE FUTURE OF PRODUCTIVITY
Andre Muscat
 
PPTX
Artificial intelligence - An Overview
Giri Dharan
 
PDF
How Does Generative AI Actually Work? (a quick semi-technical introduction to...
ssuser4edc93
 
PPTX
Generative AI Use cases for Enterprise - Second Session
Gene Leybzon
 
PPTX
Bias in Artificial Intelligence
Neelima Kumar
 
PDF
Bias in AI-systems: A multi-step approach
Eirini Ntoutsi
 
PPTX
Combatting Bias in Machine Learning
Ayodele Odubela
 
PDF
Toward Trustworthy AI
Nozha Boujemaa
 
Responsible AI in Industry (ICML 2021 Tutorial)
Krishnaram Kenthapadi
 
Generative AI Risks & Concerns
Ajitesh Kumar
 
AI and the Impact on Cybersecurity
Graham Mann
 
Responsible AI
Neo4j
 
How do we train AI to be Ethical and Unbiased?
Mark Borg
 
Exploring Opportunities in the Generative AI Value Chain.pdf
Dung Hoang
 
Algorithmic Bias: Challenges and Opportunities for AI in Healthcare
Gregory Nelson
 
What is Artificial Intelligence | Artificial Intelligence Tutorial For Beginn...
Edureka!
 
Artificial intelligence
funpathshala
 
Responsible AI
Neo4j
 
9 Examples of Artificial Intelligence in Use Today
IQVIS
 
An Introduction to Generative AI - May 18, 2023
CoriFaklaris1
 
GENERATIVE AI, THE FUTURE OF PRODUCTIVITY
Andre Muscat
 
Artificial intelligence - An Overview
Giri Dharan
 
How Does Generative AI Actually Work? (a quick semi-technical introduction to...
ssuser4edc93
 
Generative AI Use cases for Enterprise - Second Session
Gene Leybzon
 
Bias in Artificial Intelligence
Neelima Kumar
 
Bias in AI-systems: A multi-step approach
Eirini Ntoutsi
 
Combatting Bias in Machine Learning
Ayodele Odubela
 
Toward Trustworthy AI
Nozha Boujemaa
 
Ad

Similar to A Tutorial to AI Ethics - Fairness, Bias & Perception (20)

PPTX
AI Bias Oxford 2017
Dr Janet Bastiman
 
PPTX
Frontiers of Computational Journalism week 6 - Quantitative Fairness
Jonathan Stray
 
PPTX
Analyzing Bias in Data - IRE 2019
Jonathan Stray
 
PPTX
Frameworks for Algorithmic Bias
Jonathan Stray
 
PDF
Measuring Model Fairness - Stephen Hoover
PyData
 
PDF
Bias in AI
Zeydy Ortiz, Ph. D.
 
PDF
Framework for developing algorithmic fairness
journalBEEI
 
PPTX
A simple Introduction to Algorithmic Fairness
Paolo Missier
 
PDF
AIF360 - Trusted and Fair AI
Animesh Singh
 
PPTX
Ansgar rcep algorithmic_bias_july2018
Ansgar Koene
 
PPTX
Algorithmic Fairness: A Brief Introduction
AnthonyMelson
 
PDF
Using AI to Build Fair and Equitable Workplaces
Data Con LA
 
PDF
Don't blindly trust your ML System, it may change your life (Azzurra Ragone, ...
Data Driven Innovation
 
PDF
Shift AI 2020: Mission Collision: AI in the era of social justice reforms" - ...
Shift Conference
 
PDF
Fair AI
Vidhya Chandrasekaran
 
PPTX
Frontiers of Computational Journalism week 5 - Algorithmic Accountability and...
Jonathan Stray
 
PDF
Discrimination, Fairness and the Design of Robots
John Danaher
 
PPTX
Ramon van den Akker. Fairness of machine learning models an overview and prac...
Lviv Startup Club
 
PDF
Introduction to AI Ethics
Gabriele Graffieti
 
PPTX
Data ethics and machine learning: discrimination, algorithmic bias, and how t...
Data Driven Innovation
 
AI Bias Oxford 2017
Dr Janet Bastiman
 
Frontiers of Computational Journalism week 6 - Quantitative Fairness
Jonathan Stray
 
Analyzing Bias in Data - IRE 2019
Jonathan Stray
 
Frameworks for Algorithmic Bias
Jonathan Stray
 
Measuring Model Fairness - Stephen Hoover
PyData
 
Bias in AI
Zeydy Ortiz, Ph. D.
 
Framework for developing algorithmic fairness
journalBEEI
 
A simple Introduction to Algorithmic Fairness
Paolo Missier
 
AIF360 - Trusted and Fair AI
Animesh Singh
 
Ansgar rcep algorithmic_bias_july2018
Ansgar Koene
 
Algorithmic Fairness: A Brief Introduction
AnthonyMelson
 
Using AI to Build Fair and Equitable Workplaces
Data Con LA
 
Don't blindly trust your ML System, it may change your life (Azzurra Ragone, ...
Data Driven Innovation
 
Shift AI 2020: Mission Collision: AI in the era of social justice reforms" - ...
Shift Conference
 
Fair AI
Vidhya Chandrasekaran
 
Frontiers of Computational Journalism week 5 - Algorithmic Accountability and...
Jonathan Stray
 
Discrimination, Fairness and the Design of Robots
John Danaher
 
Ramon van den Akker. Fairness of machine learning models an overview and prac...
Lviv Startup Club
 
Introduction to AI Ethics
Gabriele Graffieti
 
Data ethics and machine learning: discrimination, algorithmic bias, and how t...
Data Driven Innovation
 
Ad

More from Dr. Kim (Kyllesbech Larsen) (20)

PDF
The best Spectrum, the best network, and smart investment strategies … lesson...
Dr. Kim (Kyllesbech Larsen)
 
PDF
What makes telco tick and what to expect from real 5G
Dr. Kim (Kyllesbech Larsen)
 
PDF
5G Standalone Will Deliver! - But What?
Dr. Kim (Kyllesbech Larsen)
 
PPTX
Network Automation & Autonomy - At Pace & Scale.
Dr. Kim (Kyllesbech Larsen)
 
PDF
AI for Finance
Dr. Kim (Kyllesbech Larsen)
 
PDF
Ethical Dilemmas in AI/ML-based systems
Dr. Kim (Kyllesbech Larsen)
 
PPTX
AI for Telcos
Dr. Kim (Kyllesbech Larsen)
 
PPTX
The Bot Generation ... Are Millennials excited or threatened by AI?
Dr. Kim (Kyllesbech Larsen)
 
PDF
Smart life 3.0
Dr. Kim (Kyllesbech Larsen)
 
PDF
How do we Humans feel about AI?
Dr. Kim (Kyllesbech Larsen)
 
PPTX
Towards 5G - super-agility, hyper-connectivity & networked intelligence
Dr. Kim (Kyllesbech Larsen)
 
PPTX
Zero Human Touch Networks
Dr. Kim (Kyllesbech Larsen)
 
PPTX
The Economics of 5G
Dr. Kim (Kyllesbech Larsen)
 
PPTX
Big Data @ NT - A Network Technology Perspective
Dr. Kim (Kyllesbech Larsen)
 
PPTX
Artificial intelligence - Digital Readiness.
Dr. Kim (Kyllesbech Larsen)
 
PPTX
Artificial intelligence - A Teaser to the Topic.
Dr. Kim (Kyllesbech Larsen)
 
PDF
Modern Telecommunication ... The Very Basic.
Dr. Kim (Kyllesbech Larsen)
 
PPTX
Beyond LTE Launch
Dr. Kim (Kyllesbech Larsen)
 
PPTX
Small cell networks why & when to care!
Dr. Kim (Kyllesbech Larsen)
 
PPTX
Digitized! Get Ready for the Next Wave of the Digital Society
Dr. Kim (Kyllesbech Larsen)
 
The best Spectrum, the best network, and smart investment strategies … lesson...
Dr. Kim (Kyllesbech Larsen)
 
What makes telco tick and what to expect from real 5G
Dr. Kim (Kyllesbech Larsen)
 
5G Standalone Will Deliver! - But What?
Dr. Kim (Kyllesbech Larsen)
 
Network Automation & Autonomy - At Pace & Scale.
Dr. Kim (Kyllesbech Larsen)
 
AI for Finance
Dr. Kim (Kyllesbech Larsen)
 
Ethical Dilemmas in AI/ML-based systems
Dr. Kim (Kyllesbech Larsen)
 
AI for Telcos
Dr. Kim (Kyllesbech Larsen)
 
The Bot Generation ... Are Millennials excited or threatened by AI?
Dr. Kim (Kyllesbech Larsen)
 
Smart life 3.0
Dr. Kim (Kyllesbech Larsen)
 
How do we Humans feel about AI?
Dr. Kim (Kyllesbech Larsen)
 
Towards 5G - super-agility, hyper-connectivity & networked intelligence
Dr. Kim (Kyllesbech Larsen)
 
Zero Human Touch Networks
Dr. Kim (Kyllesbech Larsen)
 
The Economics of 5G
Dr. Kim (Kyllesbech Larsen)
 
Big Data @ NT - A Network Technology Perspective
Dr. Kim (Kyllesbech Larsen)
 
Artificial intelligence - Digital Readiness.
Dr. Kim (Kyllesbech Larsen)
 
Artificial intelligence - A Teaser to the Topic.
Dr. Kim (Kyllesbech Larsen)
 
Modern Telecommunication ... The Very Basic.
Dr. Kim (Kyllesbech Larsen)
 
Beyond LTE Launch
Dr. Kim (Kyllesbech Larsen)
 
Small cell networks why & when to care!
Dr. Kim (Kyllesbech Larsen)
 
Digitized! Get Ready for the Next Wave of the Digital Society
Dr. Kim (Kyllesbech Larsen)
 

Recently uploaded (20)

PDF
Modernizing your data center with Dell and AMD
Principled Technologies
 
PDF
TokAI - TikTok AI Agent : The First AI Application That Analyzes 10,000+ Vira...
SOFTTECHHUB
 
PDF
Machine learning based COVID-19 study performance prediction
IAESIJAI
 
PPT
Teaching material agriculture food technology
LiaRayya
 
PDF
Mobile App Security Testing_ A Comprehensive Guide.pdf
flufftailshop
 
PPTX
Big Data Technologies - Introduction.pptx
kuthubussaman1
 
PDF
Encapsulation_ Review paper, used for researhc scholars
gurumoop
 
PPTX
MYSQL Presentation for SQL database connectivity
Swati270511
 
PDF
Spectral efficient network and resource selection model in 5G networks
IAESIJAI
 
PPT
“AI and Expert System Decision Support & Business Intelligence Systems”
ZubinRadhakrishnan
 
PDF
Building Integrated photovoltaic BIPV_UPV.pdf
SandeepSingh286037
 
PDF
Unlocking AI with Model Context Protocol (MCP)
Brian McKeiver
 
PPTX
PA Analog/Digital System: The Backbone of Modern Surveillance and Communication
AVTRON Technologies LLC
 
PDF
cuic standard and advanced reporting.pdf
SimoneTitaniumTomase
 
PDF
Architecting across the Boundaries of two Complex Domains - Healthcare & Tech...
Peter Tan
 
PDF
7 ChatGPT Prompts to Help You Define Your Ideal Customer Profile.pdf
SOFTTECHHUB
 
PDF
Approach and Philosophy of On baking technology
30anthuong17
 
PDF
NewMind AI Monthly Chronicles - July 2025
NewMind AI
 
PDF
Per capita expenditure prediction using model stacking based on satellite ima...
IAESIJAI
 
PDF
Reach Out and Touch Someone: Haptics and Empathic Computing
Mark Billinghurst
 
Modernizing your data center with Dell and AMD
Principled Technologies
 
TokAI - TikTok AI Agent : The First AI Application That Analyzes 10,000+ Vira...
SOFTTECHHUB
 
Machine learning based COVID-19 study performance prediction
IAESIJAI
 
Teaching material agriculture food technology
LiaRayya
 
Mobile App Security Testing_ A Comprehensive Guide.pdf
flufftailshop
 
Big Data Technologies - Introduction.pptx
kuthubussaman1
 
Encapsulation_ Review paper, used for researhc scholars
gurumoop
 
MYSQL Presentation for SQL database connectivity
Swati270511
 
Spectral efficient network and resource selection model in 5G networks
IAESIJAI
 
“AI and Expert System Decision Support & Business Intelligence Systems”
ZubinRadhakrishnan
 
Building Integrated photovoltaic BIPV_UPV.pdf
SandeepSingh286037
 
Unlocking AI with Model Context Protocol (MCP)
Brian McKeiver
 
PA Analog/Digital System: The Backbone of Modern Surveillance and Communication
AVTRON Technologies LLC
 
cuic standard and advanced reporting.pdf
SimoneTitaniumTomase
 
Architecting across the Boundaries of two Complex Domains - Healthcare & Tech...
Peter Tan
 
7 ChatGPT Prompts to Help You Define Your Ideal Customer Profile.pdf
SOFTTECHHUB
 
Approach and Philosophy of On baking technology
30anthuong17
 
NewMind AI Monthly Chronicles - July 2025
NewMind AI
 
Per capita expenditure prediction using model stacking based on satellite ima...
IAESIJAI
 
Reach Out and Touch Someone: Haptics and Empathic Computing
Mark Billinghurst
 

A Tutorial to AI Ethics - Fairness, Bias & Perception

  • 1. Ethics for Artificial Intelligent Beings. AI Ethics Workshop, 25th of June, 2018, Bonn, Germany. Dr. Kim Kyllesbech Larsen, Deutsche Telekom. Do we need to care about it? AI
  • 2. 2 Dr. Kim K. Larsen / How do we Humans feel about AI? What about Us?
  • 3. HOW DO YOU FEEL ABOUT A.I.? 16% 36% 48% Negative Neutral Positive 20% of respondents are enthusiastic about AI. 15% are uncomfortable or scared about AI. Millennials are significantly more negative towards AI. SurveyMonkey “Artificial Intelligence & Human Decision Making Sentiment Survey “ (November 2017); 467 responses. (Note: this data does not include Millennial statistics)
  • 4. DO YOU BELIEVE THAT AI COULD IMPACT YOU OR YOUR FRIENDS JOBS? Yes, 47% No, 53% Yes, 20% No, 80% Do you believe your job could be replaced by an AI? Thinking of your friends, do you believe their jobs could be replaced by an AI? SurveyMonkey “Millennials – Digital Challenges & Human Answers Survey “ (March 2018), average age 28.8 years.
  • 5. How do you think AI’s will impact your child’s or children’s future in terms of job & income? SurveyMonkey “Artificial Intelligence & Human Decision Making Sentiment Survey “ (November 2017); 467 responses. 20% 16% 64% Worse Same as Today Better 31% 23% 46% Worse Same as Today Better Men with Children above 18 yrs of Age Men with Children under 18 yrs of Age
  • 6. Note: The gender bias on this slide is completely intended.
  • 8. A DIE LIVE • AN AI-BASED AUTONOMOUS CAR SLIPS ON AN ICY MOUNTAIN ROAD OVER TO OPPOSITE SIDE OF ROAD WHERE A NORMAL CAR IS APPROACHING. • THE AI IS PROGRAMMED TO NOT DELIBERATELY CAUSE INNOCENT BYSTANDERS HARM. • THE AI CONTINUES THE CARS DIRECTION 200 METER DOWN THE ROCKY VALLEY. A FAMILY OF 4 PARISH. • “SAVING” THE 1 PASSENGER OF THE APPROACHING CAR.
  • 9. B DIE LIVE • AN AI-BASED AUTONOMOUS CAR SLIPS ON AN ICY MOUNTAIN ROAD OVER TO OPPOSITE SIDE OF ROAD WHERE A NORMAL CAR IS APPROACHING. • THE AI IS PROGRAMMED TO SAVE ITS PASSENGERS AT ALL COST FROM MORTAL DANGER IF THE LIKELIHOOD OF SUCCESS IS HIGHER THAN 50%. • THE AI COLIDES WITH THE APPROACHING CAR & SAVES ITS 4 PASSENGERS. • THE APPROACHING NORMAL CAR IS PUSHED 200 METERS DOWN INTO THE ROCKEY VALLEY & ITS 1 PASSENGER PERISH.
  • 10. YOU ARE THE AI DESIGNER! Q1: SHOULD YOU PROGRAM THE AUTONOMOUS CAR AI TO SAVE ITS DRIVER (& POSSIBLE PASSENGERS) AT ALL “COST”? Q2: WOULD YOU DEFINE A “COST” THRESHOLD?
  • 11. Immanuel Kant Jeremy Bentham  Deontological Ethics  Duty-based.  Rule-based (e.g., Rule of Law).  Asimov’s Laws.  10 Commandments.  Golden Rules.  Moral imperative.  Utilitarian Ethics  Consequentialism.  The “greater” good.  "it is the greatest happiness of the greatest number that is the measure of right and wrong.”
  • 12. An ethical framework will depend on the cultural, political & religious background of the issuer.
  • 13. AI = ? = ? = ? = ? = ?AI AI AI AI AI = ? AI For an AI Ethical Framework Design Does religious background matter? Does socio-cultural background matter?
  • 14. Societal Ethics (Public, Policy & Law Makers) Business Ethics (Corporations) Data Universe (all data cumulated) Part of the Data Universe deemed permissible by societal norms for algorithmic processing (analysis). Note that all of what might be deemed permissible from a societal perspective might not be acceptable from a business/commercial ethical perspective, Part of the Data Universe deemed permissible by business ethical norms for algorithmic processing (analysis). Note this can (should) never be larger than what is in general acceptable to society or the rule of law. Algorithm sub-space … that could be acting on the Data Universe under given societal or business ethical guidelines Part of the Data Universe that has been deemed impermissible to process algorithmically by Society incl. Businesses. E.g, due to privacy concern, un-ethical (within societal norms & laws) use, etc..
  • 15. What Ethics do you Wish for?
  • 17. KIM = WOMAN NURSE = WOMAN DOCTOR = MAN COOKING = WOMAN MANAGER = MAN HIGH RISK LOW RISK https://www.propublica.org/article/machine- bias-risk-assessments-in-criminal-sentencing “Men also like shopping” by Jieyu Zhao et al. By Joy Buolamwini et al https://newrepublic.com/article/144644/turns- algorithms-racist As AIbecomes more andmorecomplex, it can become difficult for even its own designers understand why it acts the way it does.
  • 18. Cognitive biases Statistical bias Contextual biases Bias is a disproportionate weight in favor of or against one thing, person, or group compared with another, usually in a way considered to be unfair. ℬO MO = E MO − O O: Observation. M: Model estimator of observation O. E: Expected value (long-run average) B: Bias of model of O relative to O. Hundreds of cognitive biases, e.g., Anchoring, Aavailability, Confirmation bias, Belief bias, framing effect, etc… (See: https://en.wikipedia.org/wiki/List_of_cognitive_biases) Academic bias, Experimenter bias, Educational bias, Religious bias, cultural bias, etc… (See: https://en.wikipedia.org/wiki/Bias#Contextual_biases)
  • 19. Basic on fairness & statistical bias ℬ = P Y = y X = x, G = b) − P Y = y X = x, G = w) X = x ∈ ℝn varables of direct interest i. e. , covariates for a given score S = S(x). Group G ∈ b, w of which individuals belong, e.g., b = black (African-Americans), w = white (Caucasians). Binary outcome indicator Y = y ∈ 0,1 depending on S, e.g., Y = 1 indicates individual High Risk (S ≥ 4) of re-offend. Y = 0 indicates individual Low Risk (S < 4) of re-offend. ℬ = 0 then score S is well calibrated unbiased . ℬ ≠ 0 then score S is said to be biased. Based on Chouldechova (2017) Fair prediction with disparate impact A study of bias in recidivism prediction instruments.
  • 20. Bias case study (1 of 4) Machine Bias in Risk Assessments of Criminal Sentencing US Distribution Arrests Likelihood of being arrested Population Total Total Male Population reference Caucasians 61% 57% 2.4% 3.7% 1.5% African-Americans 13% 27% 5.2% 7.9% 0.7% Others 26% 16% 1.6% 2.4% 0.4% Total 100% 100% 2.6% 3.9% 2.6% Race Arrest Likelihood within Race • In USA there are ca. 4.5× more Caucasians than African-Americans. • Still African-Americans are more than 2+ times more likely to be arrested. • Ca. 11% more African-Americans males (79%) where repeat offender compared to Caucasians (68%). • Ca. 3% more African-Americans males (18%) where violent recidivist compared to Caucasians (15%). Propublica found : a. Odds of a African-American male being assessed to have a high risk of recidivism is double that for a Caucasian male. b. Substantially higher amount (2× - 3×) of False Positives among AAs compared to Cs.
  • 21. Bias case study (2 of 4) Machine Bias in Risk Assessments of Criminal Sentencing High Risk of Recidivism differences between African-Americans & Caucasians 𝓑 The difference in getting a High Risk Score if you are a Black Male vs White Male. P(High Risk | Male, African-American) – P(High Risk | Male, Caucasian) > 0
  • 22. Bias case study (2 of 4) Machine Bias in Risk Assessments of Criminal Sentencing High Risk of Recidivism differences between African-Americans & Caucasians 𝓑 The difference in getting a High Risk Score if you are a Black Male vs White Male and being less than 25 yrs old P(High Risk | Male, Age<25yrs, African-American) – P(High Risk | Male, Age<25yrs, Caucasian) > 0
  • 23. Bias case study (2 of 4) Machine Bias in Risk Assessments of Criminal Sentencing No statistically significant difference High Risk of Recidivism differences between African-Americans & Caucasians ℬ = P Y = 1 𝐗 = 𝐱, G = African − American) − P Y = 1 𝐗 = 𝐱, G = Caucasian) 𝑌 = 1 𝑓𝑜𝑟 4 ≤ 𝑆(𝑥) ≤ 10 𝐻𝑖𝑔ℎ 𝑅𝑖𝑠𝑘 𝑜𝑓 𝑅𝑒𝑐𝑖𝑑𝑖𝑣𝑖𝑠𝑚 0 𝑓𝑜𝑟 𝑆 𝑥 < 4 𝐿𝑜𝑤 𝑅𝑖𝑠𝑘 𝑜𝑓 𝑅𝑒𝑐𝑖𝑑𝑖𝑣𝑖𝑠𝑚 X = x ∈ 𝑀𝑎𝑙𝑒, 𝐿𝑒𝑠𝑠 − 𝑡ℎ𝑎𝑛 − 25, 𝐹𝑒𝑙𝑜𝑛𝑦, 𝑃𝑟𝑖𝑜𝑟𝑠, 𝐽𝑢𝑣𝑒𝑛𝑖𝑙𝑒 𝑃𝑟𝑖𝑜𝑟𝑠, 𝑉𝑖𝑜𝑙𝑒𝑛𝑡 𝑅𝑒𝑐𝑖𝑑 𝓑
  • 24. 44 11 15 31 25 14 14 47 62 8 15 15 27 21 7 45 69 5 17 9 2 26 2 70 33 11 17 39 African-American Caucasian Males Below 25 yrs Below 25 yrs Above 25 yrs Above 25 yrs 0 23 0 77 6 33 7 54 16 14 9 69 51 6 26 17 2 26 2 70 33 11 17 39 2 26 2 70 33 11 17 39 2 Yr recid. No 2 Yr recid. 2 Yr recid. No 2 Yr recid. 2 Yr recid. No 2 Yr recid. 2 Yr recid. No 2 Yr recid. 0 14 0 86 0 29 0 71 0 13 3 84 0 0 33 67 12 42 15 30 78 0 22 0 75 0 25 0 93 0 7 0 No Juv. Priors + No Priors Juvenile Priors Juvenile Priors Juvenile Priors Juvenile Priors No Juv. Priors + No Priors No Juv. Priors + No Priors No Juv. Priors + No Priors Bias case study (3 of 4) TN LOW RISK FP HIGH RISK FN LOW RISK TP High RISK Note: the binary classification model used here does include Race to predict recidivism risk. Ignoring race in the model does not lead to substantially less biased results. FP → A Human is incorrectly (“falsely”) assessed to have a High Risk of re-offending → Gets a substantially stricter treatment. FN → A Human is incorrectly (“falsely”) assessed to have a Low Risk of re-offending → Gets a substantially lighter treatment. Might result in serious crime. Confusion Matrix (all normalized)
  • 25. Bias case study (4 of 4) Debiasing strategies (non exhaustive). 25 TN LOW RISK 14 FP HIGH RISK 14 FN LOW RISK 47 TP High RISK 62 TN LOW RISK 8 FP HIGH RISK 15 FN LOW RISK 15 TP High RISK African-American Males Caucasian Males Binary Classification Model of Low vs High Risk recidivism. Model include; Race (African-American, Caucasian & Other), Gender (female / male), age category (below & above 25), Charge Degree (misdemor & felony), #Juvenile prior count (≥0). #Prior counts (≥0), 2 Yr recidivism (0/1) & Violent Recidivism (0/1). No de-biasing 28 TN LOW RISK 10 FP HIGH RISK 19 FN LOW RISK 42 TP High RISK 58 TN LOW RISK 9 FP HIGH RISK 15 FN LOW RISK 18 TP High RISK African-American Males Caucasian Males Binary Classification Model of Low vs High Risk recidivism. Model note; Race has been taken out of the model. Remove race from model 32 TN LOW RISK 8 FP HIGH RISK 24 FN LOW RISK 36 TP High RISK 57 TN LOW RISK 5 FP HIGH RISK 22 FN LOW RISK 17 TP High RISK African-American Males Caucasian Males Binary Classification Model of Low vs High Risk recidivism. Model note; Race has been taken out of the model. The training data has been rebalanced from (CC, AA, OTH)=(34%, 51%, 15%) to the demographic blend (61%, 13%, 26%). However, it does result in less training data as the proportion of African- Americans (AA) are dramatically reduced, i.e., likely hurting accuracy. Remove race from model & re-balance race mix to demographic blend FPR = 36 FNR = 22 FPR = 11 FNR = 43 FPR = 15 FNR = 32 FPR = 9 FNR = 51 FPR = 28 FNR = 30 FPR = 17 FNR = 45
  • 26. The “Unfairness Law” On group fairness. If a model satisfy predictive parity, but the prevalence differs between groups, then that model cannot achieve equal False Positive & False Negative Rates across those groups. See: Chouldechova (2017) Fair prediction with disparate impact A study of bias in recidivism prediction instruments Base Rate P(Y=1 | G = g) Positive Predictive Value PPV = TP / (TP + FP) Not all fairness criteria can be satisfied at the same time!
  • 27. Is Google Translate misogynist? English → Hungarian → English He is a Nurse. She Is a Medical Doctor. Note: Hungarian language does not have gender-specific pronouns and lacks grammatical gender. If starting point is “The Woman is a Doctor, the Man is a Nurse”, Google translate from and back to English via Hungarian will work better. Note the Hungarian Ő has the meaning of She/He You will get the same translation bias with “He is an Assistant. She is a Manager.”
  • 28. Father is to a doctor as a Mother is to a nurse Man is to computer programmer as Woman is to homemaker Boy is to gun as Girl is to Doll Man is to Manager as Woman is to Assistant Gender biases. https://developers.google.com/machine-learning/fairness-overview/
  • 29. Man is to Manager as Woman is to Assistant Approach to de-biasing biased representations. bias non-bias Assistant Manager Man Woman He She Bolukbasi et al (2016), “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings”, 30th Conferenceon Neural Information Processing Systems (NIPS2016), Barcelona, Spain DefinitionalNeutralWords Word Embeddings e ∈ ℝ 𝑛 Softmax Word Embeddings e ∈ ℝ 𝑛 3 Bn Words 3 mn 300-dim English words w2vNEWS vector • Find bias direction based on difference between definitional opposing word vectors (e.g., he-she, male-female, …) • Neutralized non-definitional words by projecting them on to the non-bias axis.
  • 30. Man is to Manager as Woman is to Assistant Approach to de-biasing biased representations. bias non-bias Assistant Manager Man Woman He She Bolukbasi et al (2016), “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings”, 30th Conferenceon Neural Information Processing Systems (NIPS2016), Barcelona, Spain Woman is closer to Assistant than Man Woman is closer to Assistant than Man Girl Boy DefinitionalNeutralWords Word Embeddings w ∈ ℝ 𝑛 • Find bias direction based on difference between definitional opposing word vectors (e.g., he-she, male-female, …) • Neutralized non-definitional words by projecting them on to the non-bias axis. • Equalize pairs of definitional words to be equidistant to Neutral Words.
  • 31. Man is to Manager as Woman is to Assistant Approach to de-biasing biased representations. bias non-bias Assistant Manager Man Woman He She Bolukbasi et al (2016), “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings”, 30th Conferenceon Neural Information Processing Systems (NIPS2016), Barcelona, Spain Equalize pairs to be equidistant to non-biased descriptor Girl Boy DefinitionalNeutralWords Word Embeddings w ∈ ℝ 𝑛 • Find bias direction based on difference between definitional opposing word vectors (e.g., he-she, male-female, …) • Neutralized non-definitional words by projecting them on to the non-bias axis. • Equalize pairs of definitional words to be equidistant to Neutral Words.
  • 32. "Gay faces tended to be gender atypical," the researchers said. "Gay men had narrower jaws and longer noses, while lesbians had larger jaws." Wang, Y., & Kosinski, M. (in press). Deep neural networks are more accurate than humans at detecting sexual orientation from facial images. Journal of Personality and Social Psychology DEEP NEURAL NETWORKS CAN DETECT SEXUAL ORIENTATION FROM FACES
  • 33. ACCURACY = 80% What’s the likelihood you are Gay given your have been “diagnosed” Gay; P(Gay │Positive Detection) = P(G) × P( PD│G) divided by ( P(G)×P( PD│G) + P(¬G )×P( PD│ ¬G ) NOTE: This estimate given here is illustrative and possible wrong. As Wang, Y., & Kosinski has not provided other numbers than their 81% accuracy and supposedly a 23% False Negative rate, i.e., algorithm predicts a Gay man to be Straight. TRUE NEGATIVE 40% TRUE POSITIVE 40% FALSE POSITIVE 10% FALSE NEGATIVE 10% ActualClass Predicted Class Males Only GAY STRAIGHT GAYSTRAIGHT P(Gay│Positive Detection) = 2% × 40% / (2% × 40% + 98% × 10%) ≈ 8% and …. after 5 positive detections the likelihood is ≈ 73% (Don’t take this analysis too serious! … I don’t). Own guestimates FP → A Human is incorrectly (“falsely”) assessed to be Gay → Can lead to severe repercussions for the individual. FN → A Human is incorrectly (“falsely”) assessed to Straight → Unlikely to have any impact. TP → A Human is correctly (“truly”) assessed to be Gay → Can lead to severe repercussions for the individual.
  • 34. ? 64x64x3 Female German Telekom DNN Architecture e.g., 128/64/32/1 (4 Layers) Trained on 6,992+ LinkedIn pictures. TRUE POSITIVES Male Polish Vodafone FALSE NEGATIVE What’s your Gender, Nationality & Employer. How much does your face tell about you?
  • 35. What’s your Gender? MANp – MANa TRUE NEGATIVE WOMANp - WOMANa TRUE POSITIVE WOMANp - MANa FALSE POSITIVE MANp - WOMANa FALSE NEGATIVE ActualClass Predicted Class On Test Data 0.45 0.32 0.16 0.07 Predicted Class ACC = 77% PRE = 74% REC = 87%
  • 36. 36 The Cost of Machine Error. (and of Human error) 20 Million Muslims in EU. Expect < 1,300 active (Muslim) terrorists in EU* ~ 1 in 15 Thousand. (note: 5,000+ Europeans estimated to have travelled to Iraq & Syria by end of 2015) Assume a sophisticated model** gives the following: Thsd. FALSE (0) TRUE (1) FALSE (0) 19,978 9.0 TRUE (1) 0.3 1.0 Costly police work and time spend on the wrong people. Predicted Actual (*) 687 (suspects) were arrested for terrorism related-offences in 2015 (source: Europol TE-SAT 2016 report). (**) e.g., Bayesian machine learning models, Deep learning methodologies, social network analysis (e.g., social physics). STILL BOOM! ? ACCURACY = 99.99% ERROR = 0.01%
  • 37. Performance of Learning Machines. (the cost of machine error vs human error) TRUE NEGATIVE (TN) TRUE POSITIVE (TP) FALSE POSITIVE (FP) FALSE NEGATIVE (FN) Actual Class Predicted Class Precision = TP TP + FP TP TP + FN Recall = TN TN + FN Accuracy = TP + TP + FP + F1-Score = + 2 1 1 Prec. Recl.  Could be Very Costly! Could be Very Costly!    (Ex. 10%) (Ex. 77%) (Ex. 99.99%) (Ex. 18%) Note: the structure of Python sklearn.metrics confusion_matrix.is used. (Positive Predictive Value) (Sensitivity)
  • 38. Ethical AI architectures – Illustration (1 of 2). Including Bias & Fairness checks with corrections (e.g., de-biasing)
  • 39. Ethical AI architectures – Illustration (2 of 2). Vanderelst, D. and Winfield, A. (2018). An architecture for ethical robots inspired by the simulation theory of cognition. Cognitive Systems Research, 48, pp.56-66. Roboticist Alan Winfield of Bristol Robotics Laboratory in the UK built an ethical trap for a robot following Asimov’s Laws. https://www.youtube.com/watch?v=jCZDyqcxwlo 42% of the time the Robot took so long to decide action that the “Human” perished.
  • 40. 40 Dr. Kim K. Larsen / How do we Humans feel about AI? What about Us?
  • 41. 41 • 4 AI-related surveys • Paid responses. • Social media responses. • Error-margin 4% - 6%.  US population focused.  18 – 75 Yrs old.  HH Income > 75k US$.  Millennials 18 – 40 yrs old (*) Median US HH income in 2016 was 58k pa. A little more than 30% of US households earn 75k or more annually (2016).  Do weHumans trust AIs?  Acceptance of Artificial Intelligence inCorporate Decision Making.  AIEthics. AI Strategy&Policy Aistrategyblog.com
  • 42. THE GOOD & THE VERY BAD “How to train your ai chatbot to become evil in less than 24 hours”.
  • 43. Humans are increasingly dependent on digital technologies. “Today, everything & more that makes us us are replicated in the digital world.
  • 45. 10% 39% 28% 13% 10% Exclusive Frequently Sometime Rarely Never How often do you rely on news from social media ? ~50% 7% 14% 43% 22% 14% Very High High About half appears truthfull Low Very Low What is your trust level in what you read on social media ? ~80%
  • 46. Is it acceptable that your personal data, residing on social media platforms, are used for ADVERTISING? INFLUENCING ? Yes, 18% No, 82% Yes, 33% No, 67% SurveyMonkey “Millennials – Digital Challenges & Human Answers Survey “ (March 2018), average age 28.8 years.
  • 48. 4% 52% 44% Below average Average Above average ~44% Thinking of your friends, do you believe their opinions & values are negatively influenced by social media? How would you characterize your ability to detect fake news compared to your friends? 50% 40% 10% Frequently Sometimes Rarely ~50%
  • 49. 4 Dr. Kim K. Larsen / How do we Humans feel about AI? A I
  • 50. HOW DO YOU FEEL ABOUT A.I.? Millennials (18 – 38 yrs old). 31% 45% 24% Negative Neutral Positive ~24% 23% 41% 36% Negative Neutral Positive ~36% SurveyMonkey “Millennials – Digital Challenges & Human Answers Survey “ (March 2018), average age 28.8 years.
  • 51. 51 How do you feel about A.I.? (All Groups) SurveyMonkey “Artificial Intelligence & Human Decision Making Sentiment Survey “ (November 2017); 467 responses. 18% 44% 38% Negative Neutral Positive 14% 30% 57% Negative Neutral Positive 16% 44% 41% Negative Neutral Positive WOMEN WITH CHILDREN UNDER 18 YRS OF AGE 8% 27% 65% Negative Neutral Positive MEN WITH CHILDREN UNDER 18 YRS OF AGE
  • 52. Would you trust an AI with a critical corporate decision? SurveyMonkey “Artificial Intelligence & Human Decision Making Sentiment Survey “ (November 2017); 467 responses. AI “Allergy” is a Real Corporate Ailment! 19% 43% 25% 12% 0% Never Infrequently About half the time Frequently Always Would you trust a critical corporate decision made by an AI?12% 2% 9% 36% 47% 6% Never Infrequently About half the time Frequently Always Would you trust a critical corporate decision made by a fellow human expert or superior?53%
  • 53. 53Dr. Kim K. Larsen / How do we Humans feel about AI? We hold AI-s to much stricter standards than our fellow humans. Who do you trust the most with corporate decisions … Your fellow Human or your Corporate AI? 2% 12% 33% 49% 4 % Never Infrequently About half the time Frequently Always Would you trust a corporate decision made by a fellow human whos sucess rate is better than 70%?53% 11% 32% 40% 16% 1% Never Infrequently About half the time Frequently Always Would you trust a critical corporate decision made by an AI which success rate is better than 70%?17% Think a minute about how forgiving we are with the accident made by human drivers versus self-driving cars! SurveyMonkey “Artificial Intelligence & Human Decision Making Sentiment Survey “ (November 2017); 467 responses.
  • 54. Do you trust that companies using AI have your best interest in mind? ≈ 25% 75% Yes No ≈
  • 56. THANK YOU!Acknowledgement Many thanks to Viktoria Anna Laufer in particular and other colleagues who have contributed with valuable insights, discussions & comments throughout this work. Also I would like to thank my wife Eva Varadi for her patience during this work. Contact: Email: kim.larsen@telekom.hu Linkedin: www.linkedin.com/in/kimklarsen Blogs: www.aistrategyblog.com & www.techneconomyblog.com Twitter: @KimKLarsen

Editor's Notes

  • #4: Negative: “I hate it”, “It scares me” & “I am uncomfortable with it” Positive” “I am comfortable with it”, “I am enthusiastic about it” & “I love it”.
  • #5: Negative: “I hate it”, “It scares me” & “I am uncomfortable with it” Positive” “I am comfortable with it”, “I am enthusiastic about it” & “I love it”.
  • #12: Emmanual Kant 1724 to 1804 – Deontological Ethics. Jeremy Bentham 1748 to 1832 – Utilitarian Ethics.
  • #13: Bahai, Buddhist, Christian, Christian Confucian, Taoist, Hindu, Islam Judaic, Judaic, Sikh, Shinto
  • #18: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing http://gendershades.org/overview.html https://newrepublic.com/article/144644/turns-algorithms-racist https://www.aclweb.org/anthology/D17-1323 “Men also like shopping” by Jieyu Zhao et al.
  • #24: https://www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm COMPAS software to generate several scores including predictions of “Risk of Recidivism” and “Risk of Violent Recidivism.”
  • #25: https://www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm COMPAS software to generate several scores including predictions of “Risk of Recidivism” and “Risk of Violent Recidivism.”
  • #40: 1st Law: “A robot may not injure a human being or, through inaction, allow a human being to come to harm.” 2nd Law: “A robot must obey orders given to it by human beings except where such orders would conflict with the First Law.” 3rd Law: “A robot must protect its own existence, as long as such protection does not conflict with the First or Second Law.”
  • #45: https://www.forbes.com/sites/bernardmarr/2017/03/06/what-is-digital-twin-technology-and-why-is-it-so-important/#7792a4332e2a Digital twin refers to a digital replica of physical assets, processes and systems that can be used for various purposes. The digital representation provides both the elements and the dynamics of how an Internet of Things device operates and lives throughout its life cycle.
  • #47: Negative: “I hate it”, “It scares me” & “I am uncomfortable with it” Positive” “I am comfortable with it”, “I am enthusiastic about it” & “I love it”.
  • #51: Negative: “I hate it”, “It scares me” & “I am uncomfortable with it” Positive” “I am comfortable with it”, “I am enthusiastic about it” & “I love it”.
  • #52: Negative: “I hate it”, “It scares me” & “I am uncomfortable with it” Positive” “I am comfortable with it”, “I am enthusiastic about it” & “I love it”.
  • #53: Berkely Dietvorst (university of Chigaco) have addressed some of those issues with his study of algorithmic aversion. Particular as it relates to forecasting and planning. People tend to avoid algorithmic augmentation particular after seeing them fail. This is very different from how we view and treat other fellow humans.