TL;DR
This paper analyzes AI ethics curricula from 105 universities worldwide using topic modeling and pedagogical frameworks. It reveals that AI ethics education is often siloed by discipline, cognitive level, and topic, leading to a fragmented understanding of responsible AI. The authors propose moving beyond these "BAG" (Build, Assess, Govern) silos for a more holistic and interdisciplinary approach.
What Problem Does This Paper Address?
As Artificial Intelligence (AI) becomes increasingly pervasive, there's a growing demand for ethical considerations in its development and deployment. This has led to a surge in AI ethics education initiatives across universities. However, the authors identify a critical gap: a lack of systematic understanding of what is being taught, how it's being taught, and whether these approaches are truly preparing students for the complexities of responsible AI. Previous studies have often been limited in scope, focusing on specific regions, disciplines, or educational levels. This paper aims to provide a comprehensive, global analysis of AI ethics curricula to identify current pedagogical practices, their limitations, and potential areas for improvement.
Methods
The researchers employed two primary methods:
- Unsupervised Topic Modeling (Latent Dirichlet Allocation - LDA): LDA was used to automatically discover hidden thematic structures within the text of 166 AI ethics syllabi. This allowed them to identify the core topics being discussed in these courses without pre-defined categories. The identified topics were then named by experts based on the most probable words associated with each topic.
- Pedagogical Analysis (Bloom's Taxonomy and Biggs' Constructive Alignment): To understand how AI ethics is taught, the study utilized Bloom's Taxonomy, a hierarchical model for classifying learning objectives (Remember, Understand, Apply, Analyze, Evaluate, Create). They also applied Biggs' Constructive Alignment principle, which emphasizes aligning learning outcomes, teaching methods, and assessments. This framework helped them analyze the cognitive levels targeted by courses and the teaching activities employed.
Datasets
The study used a curated list of tech ethics curricula from Fiesler (2018) as an initial dataset, containing 259 courses. This was filtered to include only English-language syllabi related to AI ethics, resulting in 123 courses. An additional 43 AI ethics courses were added to create a final corpus of 166 syllabi from 105 universities worldwide.
Key Insights and Ideas
- The "BAG" Model of Siloed Education: The analysis revealed a prevalent pedagogical model, termed "BAG" (Build, Assess, Govern), where AI ethics education is often compartmentalized. "Build" typically relates to the "Build" approach, "Assess" to "Evaluate/Analyze," and "Govern" to "Understand/Remember." This suggests a tendency to teach AI ethics in isolated disciplinary, topical, and cognitive silos.
- Disciplinary Specialization: Different disciplines tend to focus on different AI ethics topics. For instance, computer science departments might lean towards "Applications" and "Data Security," while humanities might focus more on "Philosophy" and "Media & Society." This specialization, while reflecting disciplinary strengths, can lead to a fragmented understanding of AI ethics.
- Geographical Influence: The continent where a university is located can influence the discipline that delivers AI ethics education, potentially shaping the approach and focus of the curriculum.
- Misalignment Between Outcomes and Activities: A significant finding is the recurrent misalignment between the intended learning outcomes (ILOs) of AI ethics courses and the teaching modalities used to achieve them. This suggests that students may not be effectively learning what the courses aim to teach.
- Need for Interdisciplinarity: The study strongly advocates for an interdisciplinary approach to AI ethics education, arguing that responsible AI design requires a holistic understanding that transcends single disciplines.
Key Results
- The analysis identified 10 main topics within AI ethics curricula, categorized into course administration and content. The content topics included "Media & society," "Philosophy," "Responsibility," "Data security," and "Applications."
- "Responsibility" was the most frequently occurring topic in the analyzed syllabi (38.5%), followed by "Media & society" and "Data security" (both 18.5%).
- The study found that computer science departments often focus on higher Bloom cognitive levels (Apply, Analyze, Evaluate, Create), while humanities and law departments tend to emphasize lower levels (Remember, Understand).
- A significant portion of the analyzed courses (166 syllabi from 105 universities) were found to exhibit misalignments between ILOs and teaching modalities.
Limitations and Future Work
- Limitations: The authors acknowledge that their analysis is based on syllabi, which may not always perfectly reflect the actual teaching practices in a classroom. The study also relied on keywords for assessing Bloom's cognitive levels, which might not capture the full nuance of the learning objectives.
- Future Work: The paper suggests striving for a "holistic training" approach that breaks down disciplinary, topic-oriented, and Bloom cognitive level barriers. They propose a generic-to-specific hybrid teaching and learning approach that connects different communities and fosters a common understanding of AI ethics. Further research could explore the impact of these different pedagogical approaches on student learning outcomes and the development of responsible AI practitioners.
Summary updated