AI Risk and Regulation: Open-Source Release of Teaching Resources from the University of Ljubljana Workshop

The risk-ai-workshop project related to the University of Ljubljana has been open-sourced, developed and maintained by munichpavel. It provides a complete set of teaching materials for the AI Risk and Regulation workshop. Core content covers topics such as high-risk AI system identification, causal inference, and adversarial regularization, with supporting Jupyter Notebook practical exercises and an automated CI/CD release process. It aims to help learners master AI risk assessment and management methods, combining theoretical and practical value.

• Original author/maintainer: munichpavel (Paul Münch)
• Affiliated institution: University of Ljubljana
• Source platform: GitHub, original link: https://github.com/munichpavel/risk-ai-workshop
• Release date: May 26, 2026 This project is a teaching resource library focused on AI risk and regulation, which has been actually used at the University of Ljubljana.

The workshop includes four core topics:

1. Introduction to AI Risk: Basic concepts, classification frameworks, risk identification and assessment methodologies, with supporting real-case analysis exercises;
2. High-Risk AI Systems: Definitions, regulatory requirements, and compliance obligations under the EU AI Act, with exercises to assess the risk levels of AI applications;
3. Causal Inference and Discrete Geometry: Covers Simpson's paradox, d-separation, do-calculus, Average Treatment Effect (ATE), etc., with Jupyter Notebook practices (using libraries like pgmpy);
4. Future and Risks of AI: Cutting-edge topics such as adversarial regularization, discussing model robustness and defense against adversarial attacks.

• Python ecosystem integration: Jupyter Notebook (interactive environment), pgmpy (probabilistic graphical models), networkx (graph theory), graphviz (visualization), causalgraphicalmodels/dowhy/pyro (causal inference);
• Development environment: Supports venv/conda virtual environments, dependency management via requirements.txt; provides a Google Colab cloud version to lower configuration barriers;
• CI/CD process: GitHub Actions implements automated building and testing, with a custom version number rule (major version corresponds to the year, minor version distinguishes pre- and post-workshop); slides are automatically built into PDF and released via Release.

• Graded exercises: 1-star (basic, 2 points), 2-star (medium, 4 points), 3-star (advanced, 8 points), with exponential scoring to encourage challenges;
• Passing criteria: Total score ≥6 points, and at least one correct answer completed for each of the three core themes;
• Submission method: Submit via Moodle system, team collaboration is allowed, deadline is June 26, 2026 (one month after the workshop ends).

• Integration of theory and practice: Combines AI regulatory theory (e.g., EU AI Act) with code practice (causal inference algorithms) to deepen understanding of core concepts;
• Balance of cutting-edge and practical: Covers the latest legislative developments and classic statistical methods, suitable for technical personnel and policy researchers;
• Open-source and extensible: The project is open-source, allowing free use/modification; its modular structure facilitates customization; the community can contribute improvements via Issue/PR.

• Target audience: AI practitioners (needing compliance knowledge), policy researchers (AI governance), data scientists (causal inference), students of related majors (computer science, statistics, public policy, etc.);
• Prerequisite recommendations: Python basics, introduction to probability and statistics, basic machine learning concepts, basic linear algebra.

risk-ai-workshop integrates academia, policy, and engineering practice, providing an open-source model reference for AI regulation education. Its CI/CD automation and cloud support lower maintenance and usage barriers, which is of great significance for cultivating interdisciplinary AI talents. For China, although the EU AI Act provisions are not directly applicable, the risk classification ideas, causal inference methodologies, and open-source resource construction experience are of reference value.