K2SP 3500 - Decoding biases in Artificial Intelligence

The course invites students to explore the issues of discrimination in AI. The deluge of public data and the advent of deep learning techniques in the last two decades have generated a lot of hope for designing more personalized and balanced public policies. However, using machine learning algorithms to make public decisions about every citizen has not come without generating fierce criticisms from activists and citizens concerned by the way they are being computed.

In this class, we invite students to get their hands on data and code and investigate on how and why public data and state-of-the-art algorithms, far from being neutral and objective, may inherently produce discrimination toward specific populations. Conversely, we will also use state-of-the-art machine learning models to audit how they may embed systematic racial or gender biases, while still considering how they fit a larger socio-technical context. These questions are particularly pertinent and increasingly complex in the current context of the advent of generative AI models.

The objectives of the class are threefold: (i) discuss the ongoing debates on algorithmic fairness and its application to data-driven policy, (ii) learn and practice large data collection, manipulation, and main families of machine learning algorithms, (iii) create your own research design to investigate some original dataset or existing algorithm and put it to the test in a collective empirical project.

Empirical explorations of real-world cases are central in this class. For these hands-on sessions, basic Python coding skills are required: defining a function, importing an external library (we will use pandas extensively), list and dictionary manipulation. Prior coursework in statistical modelling (esp. regression analysis) is a plus but not required.

The pedagogical format is strongly oriented toward a workshop-style class. Each class session will involve a collective discussion of the readings (30mins); a short lecture on the key concepts for the session (45mins-1hr); and a hands-on coding portion (30-45mins). Some sessions may involve slightly more or less discussion, lecture, or coding time depending on the complexity of the topic.

There will be three main assessments during the semester:

• The first assignment (1/2 of the final grade) is a collective algorithmic auditing project. In groups, students will either (a) investigate an existing dataset used to evaluate AI-based systems or (b) construct a new dataset that can be used for evaluating a system of interest. In either case, the project will entail designing an experimental plan for characterizing how the dataset and AI system embed systematic biases. The project will take the form of a website visualizing and describing key results.
• The second assignment (1/4 of the final grade) is a short, reflective essay (approx. 5pgs) to be written individually. The essay should contextualize your group’s project using at least one of the readings from the semester; weigh the strengths and the weaknesses of your approach; and state what you would do differently if you were to iterate on it again.
• The third assignment (1/4 of the final grade) is to lead the paper discussion for one week of class, in a group of 4 or 5 students. Leading the discussion involves summarizing the main points of the paper and posing some critical questions of the paper to get the discussion started.

In addition to the main assignments, up to one extra point is available for active participation during the semester (asking questions, participating in discussions, etc.)

In Class Presence: 2 hours a week / 24 hours a semester

Online learning activities: 20 minutes a week / 4 hours a semester

Reading and Preparation for Class: 45 minutes a week / 18 hours a semester

Research and Preparation for Group Work: 2 hours a week / 24 hours a semester

Research and Writing for Individual Assessment: 20 minutes a week / 4 hours per semester