• Introduction to AI Ethics

    Course overview

    This course will give students the tools to understand the main ethical and legal concerns surrounding AI, empowering students to incorporate ethical and regulatory constraints in their design of AI solutions, an approach called "human-centric by design", or "ethics by design". 

    The course schedule is as follows:


    Date (Tuesdays 13:30 - 16:45)

     

    Teacher

    Key points covered

    23/11/2021

    AI ethics, fundamental rights and law

    Winston Maxwell

    ·      Difference between ethical principles and law

    ·      Fundamental rights texts affecting AI

    ·      Managing ethical/fundamental rights tradeoffs in AI

    ·      Introduction to the proposed EU AI Regulation

    30/11/2021

    Privacy and data security

    Fabian Suchanek

    ·    ensuring security, challenges of anonymization

    07/12/2021

    Privacy and data security -2

    Fabian Suchanek

    Fun hackathon

     

    14/12/2021

    Bias and Fairness 

    Sophie Chabridon 

    -Definitions

    -Individual and group fairness

    -Sources of discrimination: bias in data, bias in algorithms

    -State of the art solutions:data diversity, algorithm transparency

    -Discussion of research papers

    11/01/2022

    Bias and Fairness -2

    Sophie Chabridon


    18/01/2022

    Social impacts

    Ada Diaconescu

    social side effects: relations, society, economy, cognitive, employment. Class discussions. 

     

    25/01/2022

    AI ethics use cases

    Winston Maxwell

    Autonomous lethal weapon systems

    Facial recognition

    Social media

    Autonomous vehicles

    AI and health

    01/02/2022

    Final exam

     

     



    Over-arching AI Ethics Themes

    Below are some over-arching themes that I’ll ask you to keep in mind throughout the course. I'll ask each student to prepare an individual poster to present a use case or anecdote highlighting one of the themes below. 

    The overarching themes are the following. 

    I.               AI and the effect on work

    a.     AI replacing workforce

    b.     What is the role of work in human existence?

    c.     AI for recruiting

    d.     Amazon Mechanical Turk

    II.             AI and the surveillance state

    a.     “Surveillance capitalism” (Shoshana Zuboff) – transforming ubiquitous surveillance and data gathering into business opportunities (Facebook and Google)

    b.     Surveillance by government: predictive policing, facial recognition, algorithms to detect terrorist threats: how to draw the right balance between privacy and public security

    III.            AI and health

    a.     Individualized, predictive medicine

    b.     Epidemic (COVID) management

    c.     Neuralink

    d.     Augmented humans, transhumanism

    e.     Robot doctors

    IV.           AI and democratic institutions

    a.     AI and manipulation of populations

    b.     Fake news, polarization, and the “post-truth” era

    c.     Election manipulation, social (cyber) warfare

    d.     Freedom of expression vs censorship

    V.             AI and human dignity

    a.     Autonomous lethal weapons: the respective role of humans and machines in warfare

    b.     Robot judges – can humans be judged by a machine? (cf. Estonia robot judges experiment)

    VI.           AI and discrimination 

    a.     Racism, gender inequality, social inequalities. Does AI make societal discriminations worse? Can AI help offset human discriminations? 

    VII.          AI and the end of serendipity

    a.     What is the role of chance in our lives, careers, scientific discoveries? By reducing the role of chance, does AI harm innovation and personal development?

    b.     Can chance be a justifiable solution for ethical dilemmas such as the trolley problem? (Alexei Grinbaum)

    c.     The role of outliers (“black swans”) in human development.

    VIII.        AI and human psychology

    a.     Human machine interactions - how can AI make humans smarter (and not dumber)

    b.     Robot companions, robot ‘emotions’

    c.     Social engineering - nudges to help affect human behavior: eg « you haven’t been walking enough today… » 

    IX.            AI and safety certification

    How does machine learning change our approach to certifying safety-critical systems?

    What AI-related safety lessons can we learn from the Boeing 737 Max failures?

    X.         Can AI save humanity from itself?

    a.     AI and climate change

    b.     AI “taking control”: Isaac Asimov laws of robotics, 2001 Space Odyssey, etc. 


    The 23 Asilomar Principles

    Developed in 2017, the Asilomar principles remain today one of the best texts on AI ethics


    Research Issues

    1) Research Goal: The goal of AI research should be to create not undirectedintelligence, but beneficial intelligence.

    2) Research Funding: Investments in AI should be accompanied by funding for research on ensuring its beneficial use, including thorny questions in computer science, economics, law, ethics, and social studies, such as:

    How can we make future AI systems highly robust, so that they do what we wantwithout malfunctioning or getting hacked?

    How can we grow our prosperity through automation while maintaining people’sresources and purpose?

    How can we update our legal systems to be more fair and efficient, to keep pace with AI, and to manage the risks associated with AI?

    What set of values should AI be aligned with, and what legal and ethical statusshould it have?

    3) Science-Policy Link: There should be constructive and healthy exchange between AI researchers and policy-makers.

    4) Research Culture: A culture of cooperation, trust, and transparency should befostered among researchers and developers of AI.

    5) Race Avoidance: Teams developing AI systems should actively cooperate to avoid corner-cutting on safety standards.

    Ethics and Values

    6) Safety: AI systems should be safe and secure throughout their operationallifetime, and verifiably so where applicable and feasible.

    7) Failure Transparency: If an AI system causes harm, it should be possible to ascertain why.

    8) Judicial Transparency: Any involvement by an autonomous system in judicial decision-making should provide a satisfactory explanation auditable by a competent human authority.

    9) Responsibility: Designers and builders of advanced AI systems are stakeholders in the moral implications of their use, misuse, and actions, with a responsibility and opportunity to shape those implications.

    10) Value Alignment: Highly autonomous AI systems should be designed sothat their goals and behaviors can be assured to align with human values throughout their operation.

    11) Human Values: AI systems should be designed and operated so as to becompatible with ideals of human dignity, rights, freedoms, and cultural diversity.

    12) Personal Privacy: People should have the right to access, manage and control the data they generate, given AI systems’ power to analyze and utilizethat data.

    13) Liberty and Privacy: The application of AI to personal data must not unreasonably curtail people’s real or perceived liberty.

    14) Shared Benefit: AI technologies should benefit and empower as manypeople as possible.

    15) Shared Prosperity: The economic prosperity created by AI should be sharedbroadly, to benefit all of humanity.

    16) Human Control: Humans should choose how and whether to delegatedecisions to AI systems, to accomplish human-chosen objectives.

    17) Non-subversion: The power conferred by control of highly advanced AI systems should respect and improve, rather than subvert, the social and civicprocesses on which the health of society depends.

    18) AI Arms Race: An arms race in lethal autonomous weapons should beavoided.

    Longer-term Issues

    19) Capability Caution: There being no consensus, we should avoid strongassumptions regarding upper limits on future AI capabilities.

    20) Importance: Advanced AI could represent a profound change in the historyof life on Earth, and should be planned for and managed with commensurate care and resources.

    21) Risks: Risks posed by AI systems, especially catastrophic or existential risks, must be subject to planning and mitigation efforts commensurate with theirexpected impact.

    22) Recursive Self-Improvement: AI systems designed to recursively self-improve or self-replicate in a manner that could lead to rapidly increasing qualityor quantity must be subject to strict safety and control measures.

    23) Common Good: Superintelligence should only be developed in the service of widely shared ethical ideals, and for the benefit of all humanity rather than one state or organization.




Classes three and four: Bias and fairness