Embedding Ethics in Computer Science

Course: Probability

In this module, students are introduced to philosophical and statistical conceptions of fairness including parity and calibration, and apply these ideas to algorithms that have disparate impacts on different protected groups, including AI algorithms that predict criminal recidivism. Along the way, students consider the ethical upshot of AI and probability in other domains, including climate change and election forecasting.

Course: Algorithms

In this module, students explore barriers to applying comparison-based algorithms in the real world: incommensurable values, imperfect proxy measures, and problems with idealization and abstraction. Students also consider threats faced by people whose personal information might be included in or excluded from the data their algorithms operate on. One version of the lecture uses shortest path algorithms as a running example; an alternative version uses sorting as the example. The associated homework problems are compiled from several different problem sets across multiple versions of the course. Each of these problems includes at least a subpart that asks students to consider what other values might be at stake in the problem, how these values should be measured or approximated, how the collection of the relevant data might impact the people involved, or how idealization and abstraction in problem formulation might lead to real-world harms.

Course: Intro to Systems

This assignment is about assembly, reverse engineering, security, privacy and trust. An earlier version of the assignment by Randal Bryant & David O'Hallaron (CMU), [accessible here](http://csapp.cs.cmu.edu/public/labs.html), used the framing story that students were defusing a ‘bomb’.

Course: Intro to Systems

The assignment is the first in an introduction to systems course. It covers bits, bytes and overflow, continuing students’ introduction to bitwise and arithmetic operations. Following the saturating arithmetic problem, we added a case study analysis about the Ariane-5 rocket launch failure. This provided students with a vivid illustration of the potential consequences of overflows as well as an opportunity to reflect on their responsibilities as engineers. The starter code is the full project provided to students.

Course: Probability

This assignment uses the tools of probability theory to introduce students to _risk weighted expected utility_ models of decision making. The risk weighted expected utility framework is then used to understand decision-making under uncertainty in the context of climate change. Which of the IPCC’s forecasts should we use? Do we owe it to future people to adopt a conservative risk profile when making decisions on their behalf? The assignment also introduces normative principles for allocating responsibility for addressing climate change. Students apply these formal tools and frameworks to understanding the ethical dimensions of climate change.

Course: Algorithms

In this module, students are introduced to the distinction between disparate treatment and disparate impact in the context of hiring and protected categories. Then, they apply this distinction in an algorithms context. They also consider the relationship between disparate impact and equality of opportunity.

Course: Probability

This assignment builds on introductory knowledge of machine learning techniques, namely the naïve Bayes algorithm and logistic regression, to introduce concepts and definitions of algorithmic fairness. Students analyze sources of bias in algorithmic systems, then learn formal definitions of algorithmic fairness such as independence, separation, and fairness through awareness or unawareness. They are also introduced to notions of fairness that complicate the formal paradigms, including intersectionality and subgroup analysis, representation, and justice beyond distribution.

Course: Intro to Systems

This lab, the last of the course, asks students to discuss the case of Therac-25, a medical device that delivered lethal radiation due to a race condition.

Course: Algorithms

In this module, students return to the problem of incommensurable values and learn about moral theories that purport to tell us how to make value judgments in the face of seeming incommensurability. Students compare and contrast these theories and explore the limits to what they can tell us.

Course: Intro to Systems

This assignment is about void * and generics. We added a case study about responsible disclosure and partiality. Students read a summary of researcher Dan Kaminsky’s discovery of a DNS vulnerability and answer questions about his decisions regarding disclosure of vulnerabilities as well as their own thoughts on partiality. The starter code is the full project provided to students.

Course: Intro to Systems

When functions have assumptions, limitations or flaws, it is vital that the documentation makes those clear. Without documentation, developers don’t have the information they need to make good decisions when writing their programs. We added a documentation component to this C string assignment. Students write a manual page for the skan_token function they have implemented, learning responsible documentation practice as they go. The starter code is the full project provided to students.

Course: OS Principles

This lecture and accompanying assignments frame operating systems as the public infrastructure of computing, which thereby requires trust by system programmers, application developers, and technology users. We then introduce a framing of trust as an unquestioning attitude, identify ways trust manifests through assumption and inference, and provides examples of partly substituting the need to trust through technical and socio-technical design.

Course: OS Principles

This lecture and accompanying assignments frame operating systems consider how contextual factors in the deployment of operating systems affect trust relationships.

Course: Algorithms

In this module, students consider how social problems like homelessness and inequality differ from the sorts of problems they have learned to solve with algorithms. Students are introduced to the concept of a wicked problem, as well as various interpretations of inequality/inequity, and explore how candidate recommendation systems could be built to promote equity in hiring.

Course: Intro to HCI

The lecture covers topics associated with power relations, the use of language, standpoint and inclusion as they arise in the context of design discovery.

Course: Programming Methodologies

This lecture introduces students to the distinction between descriptive and normative language, the concepts of bias and fairness, the ethical implications of problem formulation, issues of representation in data, and the ethics of image manipulation.

Course: Reinforcement Learning

We say we want AI to be value-aligned. But what exactly does this mean? An AI agent might act differently depending on whether it is aligned to its user's intentions, its user's preferences, its user's best interests, or overall moral rightness. This lecture gets students to consider these differences and reflect on whether different targets of alignment might be more or less appropriate for different contexts (for example, LLM chatbots). It also introduces them to the basics of some moral theories and considers the question of whether these theories offer us any help in creating value-aligned AI.

Course: Intro to HCI

The lecture presents the concept of values in design. It introduced the distinction between intended and collateral values; discusses the importance of assumptions in the value encoding process; and presents three strategies to address value conflicts that arise as a result of design decisions.

Course: Intro to HCI

This assignment asks students to consider what values are encoded in their product and the decisions they make in the design process; whether there are conflicting values; and how they address existing value conflicts.

Course: Programming Methodologies

In this lecture students learn what values are and how they show up in the scientific and engineering process.

Course: AI Principles

After successfully creating a component of a self-driving car – a (virtual) sensor system that tracks other surrounding cars based on noisy sensor readings – students are prompted to reflect on ethical issues related to the creation, deployment, and policy governance of advanced technologies like self-driving cars. Students encounter classic concerns in the ethics of technology such as surveillance, ethics dumping, and dual-use technologies, and apply these concepts to the case of self-driving cars.

Course: AI Principles

In Problem 3 of this assignment, “Ethical Issue Spotting,” students explore the ethics of four different real-world scenarios using the ethics guidelines produced by a machine learning research venue, the NeurIPS conference. Students write a potential negative social impacts statement for each scenario, determining if the algorithm violates one of the sixteen guidelines listed in the NeurIPS Ethical Guidelines. In doing so, they practice spotting potential ethical concerns in real-world applications of AI and begin taking on the role of a responsible AI practitioner.

Course: Intro to HCI

This assignment asks students to evaluate their peers’ projects through a series of heuristics and to respond to others’ evaluations of their projects. By incorporating ethics questions to this evaluation, we prompt them to consider ethical aspects as part of a product’s design features which should be evaluated alongside other design aspects.

Course: Programming Methodologies

In this assignment, students write a choose-your-own-adventure game that harnesses the power of generative AI. In the end, they are asked to think about the ethical implications of using generative AI in storytelling applications.

Course: From Language to Information

Students solve a technical problem to understand how information retrieval systems rank relevant content. They practice computing precision and recall. Then, they are prompted to think about how these systems and metrics can interact with values like privacy.

Course: From Language to Information

Students solve a technical problem to understand how simple neural networks work. Then, more broadly, they are prompted to think about vulnerabilities in large language model (LLM) chatbots.

Course: Intro to HCI

This assignment asks students to consider what values are encoded in their product and the decisions they make in the design process; whether there are conflicting values; and how they address existing value conflicts.

Course: AI Principles

This assignment is about Markov Decision Processes (MDPs). In Problem 5, we use the MDP the students have created to model how a coastal city government’s mitigation choices will affect its ability to adapt to rising sea levels over the course of multiple decades. At each timestep, the government may choose to invest in infrastructure or save its surplus budget. But the amount that the sea will rise is uncertain: each choice is a risk. Students model the city’s decision-making under two different time horizons, 40 or 100 years, and with different discount factors for the well-being of future people. In both cases, they see that choosing a longer time horizon or a smaller discount factor will lead to more investment now. Students then are introduced to five ethical positions on the comparative value of current and future generations’ well being. They evaluate their modeling choices in light of their choice of ethical position.

Course: From Language to Information

Students solve a technical problem to understand the math behind the Naive Bayes sentiment classifier. Then, they are prompted to think about how the underlying mathematical assumptions of the Naive Bayes algorithm can cause ethical problems, like bias.

Course: Intro to HCI

With this assignment, students will reflect on the group of users their project is intended to serve; their reasons for selecting these users; the notion of an “extreme user;” and the reason why their perspectives are valuable for the design process. I also asks them to reflect on what accommodations they make for their interviewees.

Course: Intro to HCI

With this assignment, students are prompted to reflect on how proposed solutions to the problems they identify may exclude members of certain communities.

Course: Programming Methodologies

This lecture briefly introduces students to encryption as a powerful tool for protecting digital privacy concerns, while introducing the limitations in its ability to resolve certain privacy concerns.

Course: AI Principles

In this assignment, students explore constraint satisfaction problems (CSP) and use backtracking search to solve them. Many uses of constraint satisfaction in real-world scenarios involve assignment of resources to entities, like assigning packages to different trucks to optimize delivery. However, when the agents are people, the issue of fair division arises. In this question, students will consider the ethics of what constraints to remove in a CSP when the CSP is unsatisfiable.

Course: AI Principles

Although each of the problems in the problem set build on one another, the ethics assignment itself begins with Problem 4: Toxicity Classification and Maximum Group Loss. Toxicity classifiers are designed to assist in moderating online forums by predicting whether an online comment is toxic or not so that comments predicted to be toxic can be flagged for humans to review. Unfortunately, such models have been observed to be biased: non-toxic comments mentioning demographic identities often get misclassified as toxic (e.g., “I am a [demographic identity]”). These biases arise because toxic comments often mention and attack demographic identities, and as a result, models learn to _spuriously correlate_ toxicity with the mention of these identities. Therefore, some groups are more likely to have comments incorrectly flagged for review: their group-level loss is higher than other groups.