new global lead –ethical ai and governance massimo pellegrino · 2019. 12. 3. · technology...
TRANSCRIPT
Responsible Artificial Intelligence
Global Lead – Ethical AI and Governance
Massimo Pellegrino
Why doeasartificialintelligenceneed ethics?
Moral permissibilityof the goals
SocietalControl
Moral permissibilityof the goals
TerrorismMoral permissibility
of the goals
Chemicaland biological warfare
Moral permissibilityof the goals
Criminal operations
Moral permissibilityof the goals
The (unintentional)consequences of algorithms
The ValueAlignment problem
The (unintentional) consequences of algorithms
The inscrutability of the machine learning models
The (unintentional) consequences of algorithms
Omega Bank in Milan is about to launch a Digital Lending solution to offer short term, pre-approved small loans to its existing customer base through mobile devices
Customers will find about the loan via social media, their internet banking application, and SMS. In minutes, the customer will gain access to the Digital Lending app and is able to apply with minimal documentation.
Let's try an example… How does it work
The (unintentional) consequences of algorithms
• Lawfulness and compliance• Data Privacy
• Fairness• Security
• Adversarial attacks• API Theft
• Human Agency• Accountability• Interpretability
• Explainability
The Digital Lending solution can carries the following ethical risks:
The (unintentional) consequences of algorithms
Trust
European legislation – GDPR framework
GDPR Provision
Specifically related to profiling
Provide information
• Article 13• Article 14• Whereas 60
Data controllers must inform data subjects also of the existence of and logic involved in any automated decision-making process,including profiling
Grant access• Article 15• Whereas 63
Data controllers must provide data subjects with information regarding the existence of any automated decision-making process and the related logic involved concerning him/her
Grant objection
• Article 21• Whereas 70
Data controllers must grant data subjects the right to object if the processing, including profiling, is based on public interest or on the legitimate interest of the data controller itself
No solely automated decision
• Article 22• Whereas 71 Data controllers must not take decision based only on automated means
Perform DPIA
• Article 35• Whereas 75
Data controller must perform a DPIA where a type of processing in particular using new technologies, and taking into account the nature, scope, context and purposes of the processing, is likely to result in a high risk to the rights and freedoms of natural persons
General provisions
Apply principles Article 5 Data controllers must process personal data, also for profiling purposes, in compliance with the lawfulness, fairness, transparency,
purpose limitation, data minimisation, accuracy, storage limitation, integrity, confidentiality and accountability principles
Ask consent Article 6The processing of personal data, also for profiling purposes, is lawful only if one of the following applies: data subject’s consent, it is necessary for the performance of a contract, for compliance with a legal obligation, to protect the vital interests of the data subject, for public interests or for the data controller’s legitimate interest.
Lawfullness and Compliance concerning Data Privacy are becoming more complex
Fairness, is a social construct, a value-based choice. There are more than 30 mathematical
definitions of fairness, and when we choose one definition of fairness, we violate some aspect of
fairness captured by the other definitions. In other words, it is impossible for every decision
to be fair to all parties.
The (unintentional) consequences of algorithms
Fairness DefinitionWhat are the different fairness definitions? Which one should we used for what purpose?
Statistical Parity: people from both protected (female) and unprotected (male) groups should have equal probability of getting loans approved. Mathematically, it implies equal acceptance rates across groups.
Accepted = 8Total males = 16Acceptance Rate = 8/16= 50%
Accepted = 4Total females = 8Acceptance Rate =4/8 = 50%
Loan
Males Females
Loan
The (unintentional) consequences of algorithms
Fairness DefinitionWhat are the different fairness definitions? Which one should we used for what purpose?
Equal Opportunity: people who pay back their loans, should have an equal opportunity of getting the loan in the first place. Mathematically, it implies equal True Positive Ratesacross various groups.
True Positives = 6False Negatives = 2True Positive Rate = 6/8= 75%
True Positives = 3False Negatives = 1True Positive Rate =3/4 = 75%
Loan
Males Females
Loan
Default Non-Default
The (unintentional) consequences of algorithms
Fairness DefinitionWhat are the different fairness definitions? Which one should we used for what purpose?
Overall Accuracy: people who pay back their loans should get loans, and people who default, should not get loans. Mathematically, it implies equal overall accuracy across various groups.
True Positives = 6False Positives =2True Negatives =6False Negatives = 2
Overall Accuracy=TP+TN/Total = 12/16 = 75%
True Positives = 3False Positives =1True Negatives =3False Negatives = 1
Overall Accuracy=TP+TN/Total = 6/8 = 75%
Loan
Males Females
Loan
Default Non-Default
The (unintentional) consequences of algorithms
DatasetFairness DefinitionThe fairness definition that we agree to apply to the decision• Accuracy• Statistical parity• Equal Opportunity
Bias is estimated relative to a specific dataset• German Credit dataset• FICO dataset
Protected AttributesThe variables on which we do not want to discriminate• Gender (male/female)• Marital status
(single/married/divorced)
Bias detection is based on the decision being made, the protected attribute, fairness definition chosen and the dataset provided
DecisionThe target variable or decision under consideration
• Loan approval
• Property ownership
Security
AdversarialAttack
Hackers could trick the Digital Lending algorithm to force it to grant loans to people
with specific characteristics
APITheft
Hackers could get access to the chatbot API and provide deceitful information to the Bank’s customers
The (unintentional) consequences of algorithms
HumanAgencyShould the Digital Lending application make completely autonomous decisions?
The (unintentional) consequences of algorithms
AccountabilityIn case of malfunctioning, who is accountable for what?
The (unintentional) consequences of algorithms
InterpretabilityData controllers must inform data subjects also of the existence of and logic involved in any automated decision-making process, including profiling
• GDPRS• Article 13• Article 14
• Whereas 60
The (unintentional) consequences of algorithms
With the Digital Lending system Omega Bank wants to pursue the goals of lending money to the highest number possible of clients maximizing at the same time the probability of loan repayment.
In doing so, a primary objective of the bank is to treat its clients with fairness, especially with regard to gender discrimination.To achieve this goals, a Statistical parity algorithm is used, which means that males and females have equal probability of getting loans approved (for more information on how the system works visit http://xxxxxxxxxx)
Interpretability example
The (unintentional) consequences of algorithms
The Digital Lending system is GDPR compliant and the current version has been tested with a xx% accuracy (for more information on how the system was developed and how it is maintained visit http://xxxxxxxxxx)
What can companiesand organizations do to mitigateethical AI risks?
1. Define an ethicalAI Code of Conduct
Apply Ethics & Legal throughout, helping clients understand the ethical implications of their use of AI
Ethical Approach
Fundamental Human Rights
Epistemic & General
Principles of AI Ethics
Contextualisation
Methodology
Operationalise Ethics through Computational
Ethics
AI Ethics Framework
PwC’s Ethical AI FrameworkEpistemic and General Ethical AI Principles
• Interpretability
• Reliability, Robustness and Security
Epistemic Principles
Set of principles that are objective and do not prescribe any specific behaviour consistent with a specific rule or value. Rather, they represent the indispensable conditions of knowledge without which it is impossible to determine whether an AIS is consistent with an ethical principle.
• Beneficial AI
• Fairness
• Human Agency
• Safety
• Data Privacy
• Lawfulness & Compliance
• Accountability
General Principles
Set of behavioural principles that have a universal scope. These principles, when contextualised, must stand valid in any possible cultural and geographical application.
2. Define an ethicalAI Governance Framework
Ensure end-to-end Governance
Model Level
1. S
trateg
y
3. Ecosystem
5. Deployment
6. Operate and Monitor
Corporate Strategy
Industry Standards & Regulations
Internal Policies & Practices
Operational Support
Compliance
Portfolio Management
Program Oversight
Delivery Approach
Technology Roadmap
Sourcing
2. P
lan
nin
g
Transition & Execution
Ongoing monitoring
Data Extraction
4. D
evel
op
men
t
Change Management
Evaluation & Check-in
Model Integration
Solution Design
Business & Data Understanding
Pre-Processing
Model Building
Who Benefits:• C-Suite• Process Owners• Risk & Compliance• Data Scientists• Consumers• Regulators
3. Identify the software tools necessary for an effective and consistent implementation and monitoring of the ethical AI principles
• Ethical AI Governance Platform• Bias detection and avoidance
applications• Explainability applications• Applications to manage data
privacy preferences
Someexamples
4. Communicate with external and internal stakeholders to find operational solutions to ethical tensions and trade-offs
"So the rate of improvement is really dramatic. We have to figure out some way to ensure that the advent of
digital super intelligence is one which is symbiotic with humanity. I think that is the single biggest existential
crisis that we face and the most pressing one."Elon Musk – CEO of Tesla, SpaceX, and OpenAI
Thank you!