Teaching Fairness, Accountability, Confidentiality, and Transparency in Artificial Intelligence through the Lens of Reproducibility

Ana Lucic, Maurits J. R. Bleeker, Sami Jullien, Samarth Bhargav, Maarten de Rijke

TL;DR

This paper details a technical, graduate-level course that teaches Fairness, Accountability, Confidentiality, and Transparency (FACT-AI) by tasking students with reproducing state-of-the-art research. By moving beyond theory to hands-on implementation, students gain a critical understanding of how ethical principles are operationalized—and often fail—in practice.

What Problem Does This Paper Address?

As AI systems increasingly influence daily life, there is a growing need for developers to understand the ethical ramifications of their work. However, "tech ethics" is often taught as a philosophical elective, which technical students may perceive as a distraction from "real" engineering.

Furthermore, the AI community is grappling with a reproducibility crisis. If a decision-making algorithm’s results cannot be replicated, it is impossible to verify if it truly adheres to fairness or transparency standards. The authors address these dual challenges by integrating FACT-AI education with the rigorous practice of scientific reproduction, ensuring that responsible AI is treated as a core technical competency rather than an afterthought.

Methods

The authors describe a mandatory, four-week intensive course for MSc AI students at the University of Amsterdam. The pedagogical approach centers on three pillars:

  1. Technical Lectures: High-level overviews of the four FACT-AI pillars and the methodology of reproducibility.
  2. Paper "Dissections": Discussion sessions where students critically analyze seminal papers. Instructors lead "dissections" to show that there is no single way to evaluate research.
  3. The Reproducibility Project (Core): Groups of 3–4 students select a recent FACT-AI paper from a top-tier conference (e.g., NeurIPS, ICML). They must re-implement the algorithm from scratch (or extend it if code is available), re-run experiments to verify claims, and write a formal reproducibility report.

The course is strategically placed in the curriculum after students have completed foundational machine learning and deep learning courses, as shown in the program structure below:

Table 1: The first year of the MSc AI program at the Univer- sity of Amsterdam.

Datasets

While the paper describes a course rather than a new model, the "datasets" involved are the peer-reviewed papers selected for reproduction. The authors curated a list of 10–15 papers based on specific criteria:

  • The topic must be FACT-AI related.
  • At least one dataset used in the original paper must be publicly available.
  • Experiments must be runnable on a single GPU (provided by the university).
  • The implementation must be feasible for a group of master's students within one month.

Key Insights and Ideas

  • Reproducibility as a Gateway to Ethics: The authors argue that reproducibility is the cornerstone of FACT-AI. You cannot claim a system is "fair" or "transparent" if its results are a "black box" that cannot be replicated by others.
  • The "Aha!" Moment of Failure: Students often found that reproducing results was significantly harder than expected. Seeing how (poorly) some research is performed was a major eye-opener, fostering a healthy skepticism toward "state-of-the-art" claims.
  • Real-World Impact as Motivation: In the first year, students contributed to an open-source repository. In the second, they submitted to the Machine Learning Reproducibility Challenge (MLRC). This gave students a goal beyond a grade: contributing to the global scientific community.
  • Technical Depth Matters: By keeping the focus technical, the course avoids the "distraction" trap. Students learn about algorithmic harm by trying to fix it in code.

Key Results

The course yielded impressive quantitative and qualitative outcomes:

  • High Student Satisfaction: 67.8% of students were (very) satisfied with the course, with over 90% satisfied with the level of the final report and workload.
  • Scientific Contribution: In the 2020–2021 iteration, 30 out of 32 groups submitted to the MLRC. 9 of these reports were accepted for publication in the ReScience journal.
  • Institutional Success: Nearly 40% of all accepted papers in the 2021 MLRC came from this single course at the University of Amsterdam.
  • Skill Development: Qualitative feedback showed students developed a "critical view" of research, stating they would be "more critical from now on, as many papers are not reproducible."

Limitations and Future Work

  • The "Online" Hurdle: Teaching during the COVID-19 pandemic made group formation and spontaneous discussion difficult. The authors noted that some online "subrooms" struggled to get conversations going.
  • Broadening Perspectives: While the technical focus was successful, the authors acknowledge the need to incorporate more non-technical perspectives, such as legal views on accountability or social science frameworks.
  • Logistical Polish: First-time authors often made minor mistakes (e.g., forgetting double-blind submission rules). Future iterations will include more explicit guidance on the standard procedures of academic publishing.
  • Future Integration: The authors plan to include "broader impact statements" in student reports to further encourage reflection on the societal consequences of AI research.

Summary updated

Abstract

In this work, we explain the setup for a technical, graduate-level course on Fairness, Accountability, Confidentiality, and Transparency in Artificial Intelligence (FACT-AI) at the University of Amsterdam, which teaches FACT-AI concepts through the lens of reproducibility. The focal point of the course is a group project based on reproducing existing FACT-AI algorithms from top AI conferences and writing a corresponding report. In the first iteration of the course, we created an open source repository with the code implementations from the group projects. In the second iteration, we encouraged students to submit their group projects to the Machine Learning Reproducibility Challenge, resulting in 9 reports from our course being accepted for publication in the ReScience journal. We reflect on our experience teaching the course over two years, where one year coincided with a global pandemic, and propose guidelines for teaching FACT-AI through reproducibility in graduate-level AI study programs. We hope this can be a useful resource for instructors who want to set up similar courses in the future.
BibTeX
@inproceedings{lucic2022teaching,
  author = {Ana Lucic and Maurits J. R. Bleeker and Sami Jullien and Samarth Bhargav and Maarten de Rijke},
  title = {{Teaching Fairness, Accountability, Confidentiality, and Transparency in Artificial Intelligence through the Lens of Reproducibility}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21558},
}
Figure