Introduction
As Natural Language Processing (NLP) systems transition from controlled research environments to ubiquitous real-world applications, the potential for societal harm—ranging from algorithmic bias and privacy violations to the spread of misinformation—has increased exponentially. While the research community has introduced various "best practices," such as data statements and model cards, educational curricula often fail to keep pace. Ethics is frequently relegated to a single guest lecture at the conclusion of a technical track, treated as an afterthought rather than a core competency.
This paper presents a blueprint for a hands-on "Ethics in NLP" course designed to bridge this gap. By moving beyond passive lectures into active, "learning by teaching" pedagogy, the authors demonstrate how to transform abstract ethical concepts into a tangible sense of professional responsibility for future AI practitioners.
Pedagogical Framework: The Protégé Effect
The course structure, refined over four years at the University of Groningen, the University of Turin, and the University of Pavia, is built on the foundation of active learning. Specifically, it leverages the "protégé effect"—a psychological phenomenon where students achieve deeper mastery of a subject when they are tasked with teaching it to others.
The curriculum is designed as a 6-week module (yielding 5-6 ECTS) that integrates theoretical grounding with practical application. The course components include:
- Weekly Lectures & Seminars: Two-hour sessions covering the entire NLP pipeline, from data collection and annotation to evaluation and reporting.
- Hands-on Labs: Practical activities designed to challenge assumptions. For instance, a "live transcription" exercise demonstrates that data is a human-made artifact rather than a neutral "given." Students also engage in first-hand annotation tasks to experience the inherent subjectivity of "gold standard" data.
- Weekly Assignments: Mandatory, non-graded tasks such as auditing ACL ethics statements or debunking over-hyped AI claims for a lay audience.
The Final Project: From Theory to Outreach
The most innovative aspect of the course is the final project, which evolved through three distinct formats to maximize student engagement and social impact:
- Interviews with Experts: Students interviewed members of the ACL Ethics Committee to understand the practical challenges of policy-making.
- High School Presentations: Students translated complex NLP ethics concepts for teenage audiences, forcing them to simplify and clarify their own understanding.
- Educational Outreach Products: Students created reusable materials—such as games, podcasts, and books—tailored for specific target audiences.
Key Insights: Data is "Made," Not "Found"
A central takeaway from the course is the realization that data is never truly neutral. Through transcription exercises, students discovered that even basic choices regarding punctuation, line breaks, and spelling involve human judgment. This shift in perspective—from viewing data as a raw resource to seeing it as a curated artifact—is crucial for developing responsible AI.
Furthermore, the course emphasizes ethics as science communication. The authors argue that a practitioner’s responsibility includes the ability to communicate risks to a broader audience. Moving from "black and white" technical metrics to nuanced social impact assessments is a critical skill for the modern engineer. To support this, the course utilizes a diverse range of materials beyond academic papers, including investigative journalism, documentaries (e.g., Coded Bias), and podcasts, to keep the content relevant and tangible.
Results and Student Impact
The course successfully produced a suite of reusable, student-made educational products that demonstrate a high level of conceptual integration:
- Card Games: Quartet (focused on privacy and bias), Ethica ex Machina (a debate-triggering game), and Debatable (a rhetoric-focused party game).
- Multimedia: Luna en Chat (an illustrated children's book about chatbots), podcasts on corporate data usage, and interactive web demos that simulate the ethical trade-offs involved in chatbot development.
Qualitative feedback indicates a significant shift in student perspectives. Testimonials highlight that the course "condensed abstract ethical thinking into crucial practical advice" and had a "concrete impact" on how students perceive their technical studies.
Limitations and Future Directions
While successful, the authors note several challenges in implementing this model:
- Scalability: The "Interview with Experts" model relies heavily on the instructors' personal networks, which may be difficult to replicate.
- Logistics: In intensive "crash course" formats, organizing external school visits can be logistically prohibitive, necessitating "mock" presentations instead.
- Interdisciplinary Gaps: Students from less technical backgrounds (e.g., Linguistics) often require additional "bridge" lectures to fully grasp the technical mechanics of language models before they can evaluate their ethical implications.
The authors hope this blueprint inspires other institutions to integrate social impact as a core component of the NLP curriculum, fostering a generation of practitioners who view ethics not as a hurdle, but as a fundamental aspect of scientific excellence.
Summary updated