Ethical Design And Decision in Autonomous And Intelligent Systems - - PowerPoint PPT Presentation

Ethical Design And Decision in Autonomous And Intelligent Systems

23/11/2017 1

Raja Cha-la

Ins-tute of intelligent Systems and Robo-cs (ISIR) University Pierre and Marie Curie, Paris

Raja.Cha-la@isir.upmc.fr Chair, The IEEE Global Ini.a.ve Ini.a.ve on Ethics of Autonomous and Intelligent Systems

Booming Applica-ons Robo-cs and AI Manufacturing, Transporta1on, Logis1cs, Agriculture, Mining, Construc1on, Health, Jus1ce, Banking, Personal services, Leisure, Defense, Interven1on, etc.

• To replace humans
• To assist and serve

humans

• To rehabilitate/augment humans

2

Clinique du Risque de l’UPMC - PSL

A Few Ethical, Legal, Social Issues Raised by A/IS

• Impact on jobs
• Personal data, privacy, intrusion, surveillance
• Transparency, explainability of algorithmic decisions
• Autonomous/learned machine decisions
• Cogni-ve and affec-ve bonds with robots
• Human dignity, integrity and autonomy
• Transforma-on and augmenta-on of humans
• Anthropomorphism and Human iden-ty
• Legal accountability and responsibility of robots
• Status of robots in the human society
• Specific robot applica-ons and usage (AWS, Sexbots)
• Fears about General/Super AI …
• …

3 23/11/2017

Ethical Concerns

• (Un)Ethical usage of robots, AI and

autonomous systems.

• Ethics of machine decisions.
• Ethically aligned design: Ethics in research,

and engineering.

23/11/2017 4

The IEEE Global Ini1a1ve on Ethics of Autonomous and Intelligent Systems

• Launched April 2016
• Mission: To ensure every stakeholder involved in the design and

development of AIS is educated, trained, and empowered to priori-ze ethical considera-ons so that these technologies are advanced for the benefit of humanity.

• Brings together mul-ple and diverse voices from academia, industry and
• rganiza-ons in the A/IS and ELS communi-es and landscapes to iden-fy and

find consensus about ELS in the development and deployment of A/IS.

• Version 1 of Ethically Aligned Design: December 2016.
• Version 2 featuring five new sec-on: December 2017. Final version by 2019.
• 11 standards proposals under discussion/development within IEEE-SA by Ad-

Hoc working groups.

5 23/11/2017

Ethically Aligned Design

A Vision for Prioritizing Human Wellbeing with Artificial Intelligence and Autonomous Systems

6

Version 1

• Released December, 2016 as a Creative Commons

doc / RFI for public input

• Created by over 100 Global AI/Ethics experts, in a

bottom up, globally open and transparent process

• Eight Committees / Sections
• Contains over eighty key Issues and Candidate

Recommendations

• Designed as the “go-to” resource to help

technologists and policy makers prioritize ethical considerations in AI/AS

23/11/2017

Ethically Aligned Design

A Vision for Prioritizing Human Wellbeing with Artificial Intelligence and Autonomous Systems

7

Version 2

• Launching December 2017 as a Creative Commons doc /

RFI for second round of public input

• Created by over 250 Global AI/Ethics experts, in a bottom

up, transparent, open and increasingly globally inclusive process

• Will incorporate over 200 pages of feedback from public

RFI and new Working Groups from China, Japan, Korea and Brazil

• Thirteen Committees / Sections
• Will contains over one hundred twenty key Issues and

Candidate Recommendations

• Designed as the “go-to” resource to help technologists and

policy makers prioritize ethical considerations in AI/AS

Cover Design TBD 23/11/2017

Current CommiMees

• General Principles
• Personal Data and Individual Access Control
• Embedding Values Into Autonomous Intelligent Systems
• Methodologies to Guide Ethical Research and Design
• Safety and Beneficence of Ar-ficial General Intelligence (AGI) and Ar-ficial Superintelligence (ASI)
• Reframing Autonomous Weapons Systems
• Economics/Humanitarian Issues
• Law
• Affec-ve Compu-ng
• Classical Ethics in Informa-on & Communica-on Technologies
• Policy
• Mixed Reality
• Wellbeing

8 23/11/2017

IEEE-SA Standards Projects for Ethically Aligned Design

9

• IEEE P7000: Model Process for Addressing Ethical Concerns During System Design
• IEEE P7001: Transparency of Autonomous Systems
• IEEE P7002: Data Privacy Process
• IEEE P7003: Algorithmic Bias Considerations
• IEEE P7004: Standard on Child and Student Data Governance
• IEEE P7005: Standard on Employer Data Governance
• IEEE P7006: Standard on Personal Data AI Agent
• IEEE P7007:

Ontological Standard for Ethically Driven Robotics and Automation Systems

• IEEE P7008: Standard for Ethically Driven Nudging for Robotic, Intelligent and

Autonomous Systems

• IEEE P7009: Standard for Fail-Safe Design of Autonomous and Semi-Autonomous

Systems

• IEEE P7010 : Wellbeing Metrics Standard for Ethical Artificial Intelligence and

Autonomous Systems

23/11/2017

Embedding Values into Autonomous Systems

10

Example Issues:

• Values to be embedded in A/IS are not all universal;

some are specific to user communities and to tasks.

• Moral overload: A/IS are usually subject to a

multiplicity of norms and values that may conflict with each other.

• A/IS can have built-in data or algorithmic biases that

disadvantage members of certain groups.

23/11/2017

Autonomy And Decision-making

11 23/11/2017

What is Autonomy?

• Autonomy: ability of an agent to determine and achieve its ac-ons by its own
• means. Rela-ve to environment and tasks. Related to adapta-on capacity.
• Opera-onal autonomy vs. decisional autonomy
• Aeainable Autonomy is rela-ve to task and environment complexity

Complexity of environment Complexity

• f task

Attainable Autonomy

Complexity

• f task

Attainable Autonomy

Complexity of environment

12 23/11/2017

Examples of Autonomy

Teleopera-on: human control Advanced automa-c control

13

Opera-onal Autonomy Opera-onal/some decisional Autonomy

23/11/2017

Opera-onal Autonomy

Da Vinci Roomba Crusher, CMU Naviga-on, LAAS

Automated Driving

14 23/11/2017

Google

Automated Driving: Usual Situa1ons and Moral Dilemmas

15 23/11/2017

• Retaking human control from a self driving car ~ 10’’

At 36 km/h the car would have moved an addi-onal 10m on its own.

23/11/2017 16

UK Department of Transport

Machine decisions

• Machine decisions are based on knowledge and

ac-on possibili-es.

• Knowledge about the environment is acquired by

sensing and through contextual informa-on.

• Knowledge is prone to be par-al and uncertain.
• Situa-ons are dynamic.
• Ac-ons might have unexpected outcomes.

17 23/11/2017

What does it mean for a machine to make ethical decisions?

• Ethical decisions are related to human dignity and well-being
• Abstract concepts such as dignity cannot be explicitly described, taught

to/understood by machines

• Machines are not autonomous as humans are because they cannot

decide their purpose and their own goals. Therefore machines cannot determine ethical values

• Therefore machines cannot make ethical decisions, but can perform

ac-ons with ethical consequences.

• Machines can only select within a bounded set of categories or decisions

provided to them directly or indirectly (e.g., learning) by a human programmer

• Accountability remains on the human who programmed the machine

23/11/2017 18

Automated Driving: Usual Situa1ons and Moral Dilemmas

19 23/11/2017

Decision under Uncertainty

• Situa-on assessment. Iden-fica-on of current state

and es-ma-on of future states {S}: Pr(s)

• Possible decisions: state-dependent ac-ons {A}
• Uncertain state transi-ons Pr(s,a,s’).
• Each ac-on and resul-ng state characterized by a

value H(a,s) reflec-ng the es-mated incurred harm.

23/11/2017 20

Ethical Approaches

• Virtue Ethics: promo-ng the personal virtues and the “good life”

(Plato, Aristotle). Agents must be virtuous for their decisions to be good

• Deon-c Ethics: Obey a moral impera-ve in all circumstances (I. Kant 1797)
• Consequen-alism, U-litarianism: “The greatest good for the greatest number”

(J. Bentham 1789, J. S. Mill 1861)

• Casuis-c approach
• Theory of Jus-ce; protect the most vulnerable (J. Rawls 1971)

21 23/11/2017

Theory of Jus1ce

• Jus-ce as fairness
• The “veil of ignorance”: To choose a decision-making system

while ignoring how you will be affected by its decisions

• Minimize harm for the most vulnerable

22

John Rawls 1921-2002

23/11/2017

23/11/2017 23

Best decision in state s: Π*(s) = argmina Pr(s, a, s’) H(a,s’) “Rawlsian” decision: to cause the least harm to the most vulnerable State characterized by a vulnerability measure according to predefined categories and to actual situa-on interpreta-on

Conclusions Related to Autonomy

• Moral choices and legal liability
• Stakeholder consensus
• Unbiased, transparent and traceable decision-making

processes

• Explainable decisions
• Robot opera-on obedient to fundamental ethical

principles, law and interna-onal humanitarian law.

• Rawlsian perspec-ve: more contextual; reduce highest

degree of harm for the most vulnerable?

• Disputable Idea of a “electronic personality” for

autonomous robots

• Accountability must remain with the humans behind the

system.

24 23/11/2017

Issues related to social and companion robots : How close to us should robots be?

• Respect privacy, in-macy
• Frame cogni-ve and affec-ve bonds with robots
• Issue with expression of emo-ons: Consider

impact on development of children’s emo-onal capaci-es

• Respect Human dignity, integrity and autonomy

23/11/2017 25

Issues: Should be we develop human-like robots?

• Human iden-ty and anthropomorphism: projec-on of humanity
• n machines?

Robots should be dis-nguishable as such.

• Robots’ in the human society: equal to humans? rights and du-es?

Respect human iden-ty and dignity.

• Specific robot applica-ons and usage (e.g., sexbots)?

How to frame ques-onable applica-ons

26