Introduction
As artificial intelligence and machine learning systems transition from controlled research environments to real-world applications, their societal impact has become impossible to ignore. From biased facial recognition to algorithms that penalize specific demographics, the "unwitting" encoding of bias is a systemic challenge. Researchers from Manhattan College and UMass Lowell addressed this by integrating a one-week ethics module into a traditional AI course, using autonomous vehicles (AVs) as a primary case study. The goal was to equip computer science students with the philosophical frameworks necessary to navigate moral complexities, such as "least-worst" outcomes in unavoidable accidents.
Methodology
The study implemented a targeted pedagogical intervention consisting of three main components:
- Interdisciplinary Lectures: A collaboration between Philosophy and Computer Science professors introduced Utilitarianism—the theory that the right action maximizes overall "utility." This framework was chosen for its compatibility with the decision-theoretic approach central to AI, where agents maximize expected value functions.
- The "Trolley Problem" Test Cases: Student pairs navigated eight scenarios inspired by the "Moral Machine" project. These cases forced choices between lives based on factors like age, number of individuals, or legal compliance (e.g., wearing a seatbelt).
- Integrated Assessment: Ethical reasoning was embedded into the curriculum, requiring a dedicated section in final project papers and a question on the final exam.
Qualitative Analysis and Student Reasoning
The study collected qualitative data from 20 student pairs across eight ethical test cases. A key finding was the "Utilitarian Trap": students naturally gravitated toward Utilitarianism because it mirrors the mathematical optimization they already practice. However, friction arose in "edge cases" where pure utility conflicted with notions of justice. For instance, in a scenario where a car must choose between hitting a passenger wearing a seatbelt or one who is not, 80% of student pairs chose to save the seatbelt wearer, shifting their reasoning from maximizing lives to not "punishing" law-abiding behavior.
Students also grappled with the encoding of value judgments. By comparing scenarios involving different demographics (e.g., a school bus vs. a casino bus), students realized that their algorithms would inevitably assign "worth" to human lives, making the "unwitting" encoding of bias a visible, technical reality.
Results and Impact
The intervention demonstrated high efficacy:
- Retention: 98% of students (62 out of 63) successfully explained and diagrammed the Trolley Problem on their final exam.
- Nuanced Application: In final projects, students moved beyond technical feasibility to ethical viability. One group simulating search-and-rescue questioned the ethics of an AI "exploring" (and potentially failing a rescue) to improve future performance in high-stakes environments.
- Critical Thinking: While students gained fluency in Utilitarianism, the authors noted the difficulty students had in critiquing the framework. The strong alignment between AI's mathematical foundations and Utilitarianism can reinforce a narrow mindset if not balanced with other ethical theories.
Conclusion
This study proves that a one-week "sprint" is sufficient to introduce basic ethical fluency in an AI course. However, future work must address the "utilitarian default" by introducing alternative frameworks like Deontology (duty-based ethics) and requiring students to formalize ethical choices in code. This would highlight where mathematical models fail to capture complex human values like justice and rights, ensuring future engineers are better prepared for the societal responsibilities of their work.
Summary updated