Applied Ethical and Governance Challenges in AI (Spring 2019)

Taught by Jonathan Zittrain and Joi Ito

John Bowers (Harvard Law School, Berkman Klein Center)

This course will pursue a cross-disciplinary investigation of the development and deployment of the opaque complex adaptive systems that are increasingly in public and private use. We will explore the proliferation of algorithmic decisionmaking, autonomous systems, and machine learning and explanation; the search for balance between regulation and innovation; and the effects of AI on the dissemination of information, along with questions related to individual rights, discrimination, and architectures of control.

The structure of the course is somewhat nontraditional (see the “Schedule Overview” section) – class days are organized as steps in an analytical process rather than by topic. We will commit three class sessions each to three such steps – diagnosis, prognosis, and intervention. Our sessions focused on diagnosis will seek to identify the mechanisms at the root of the problems we examine, sessions focused on prognosis will explore the problems’ implications for society at large, and sessions focused on intervention will consider the costs and benefits of prospective solution paths. This process will be bound together by a central theme – tradeoffs in the design of AI systems. Fixing AI’s problems will mean imposing burdens and constraints on the performance of critical systems, but the costs of sticking with the status quo might be much worse.

We will apply this analytical process to 3 distinct problem areas, committing one diagnosis session, one prognosis session, and one intervention session to each. Our three problem areas are fairness, interpretability, and adversarial example attacks. The first two are well developed within the AI literature and have been rearing their ugly heads in the real world for years, but are still host to numerous unresolved questions. The last is still largely terra incognita – adversarial examples haven’t yet made it out of the lab, but there are strong indications that they may do so soon. The literature around adversarial examples is highly technical, and needs much more attention from lawyers and social scientists.

2 Part I: Diagnosis 2 Part I: Diagnosis

2.1 Diagnosing problems of fairness 2.1 Diagnosing problems of fairness

2.1.1. [COPIES TO BE DISTRIBUTED] Weapons of Math Destruction by Cathy O’Neil, Broadway Books (2016). Read Introduction and Chapter 1: “Bomb Parts: What Is a Model?”

[COPIES TO BE DISTRIBUTED] Weapons of Math Destruction by Cathy O’Neil, Broadway Books (2016). Read Introduction and Chapter 1: “Bomb Parts: What Is a Model?”

2.1.2. [OPTIONAL] "The scored society: due process for automated predictions" by Danielle Keats Citron and Frank Pasquale, Washington Law Review (2014)

[OPTIONAL] "The scored society: due process for automated predictions" by Danielle Keats Citron and Frank Pasquale, Washington Law Review (2014)

2.2 Diagnosing problems of interpretability 2.2 Diagnosing problems of interpretability

2.2.1. “The Mythos of Model Interpretability” by Zachary C. Lipton, ArXiv (2016)

“The Mythos of Model Interpretability” by Zachary C. Lipton, ArXiv (2016)

2.2.2. [OPTIONAL] “Towards a rigorous Science of Interpretable Machine Learning” by Finale Doshi-Velez and Been Kim, ArXiv (2017)

[OPTIONAL] “Towards a rigorous Science of Interpretable Machine Learning” by Finale Doshi-Velez and Been Kim, ArXiv (2017)

2.3 Diagnosing vulnerabilities to adversarial examples 2.3 Diagnosing vulnerabilities to adversarial examples

2.3.1. “Motivating the Rules of the Game for Adversarial Example Research” by Justin Gilmer et al., ArXiv (2018).

“Motivating the Rules of the Game for Adversarial Example Research” by Justin Gilmer et al., ArXiv (2018).

2.3.2. [RECOMMENDED] “Intriguing properties of neural networks” by Christian Szegedy et al., ArXiv (2013)