AI Safety in Practice

📘 What’s Covered

The workbook is part of the broader AI Ethics and Governance in Practice curriculum and is designed primarily for civil servants engaged in AI projects. It focuses on four core AI safety objectives—performance, reliability, security, and robustness—and translates them into actionable practices for public sector AI teams.

Part 1: Key Concepts provides a rich conceptual foundation:

Performance is explained through metrics like accuracy, precision, recall, F1-score, and AUC-ROC.
Reliability addresses model drift, brittleness, overfitting, and unpredictability.
Security includes data poisoning, model inversion, and adversarial attacks.
Robustness covers model behavior under adversarial conditions and transfer learning vulnerabilities.

Part 2: Practical Implementation offers a stage-by-stage guide across the AI lifecycle—from planning and problem formulation to deployment and deprovisioning. It introduces a Safety Self-Assessment and Risk Management Template, which is meant to be used iteratively across project phases.

The workbook also features workshop-ready activities, such as:

Conceptualizing AI safety terms with teams,
Identifying risks using public sector case studies,
Mapping those risks to specific lifecycle stages.

Visual tools—like the safety wheel on page 30—help teams embed safety into design, development, and deployment phases.

This facilitator version includes rich annotations for delivering the workshop effectively, with clear co-facilitation tips and reflective discussion questions throughout.

💡 Why it matters?

AI safety is often abstract or sidelined. This workbook breaks that trend by showing how safety can be applied methodically, not just as a checklist but as an evolving mindset baked into every decision. It’s especially valuable for public institutions where risk tolerance is low and accountability is high. The mix of technical depth and team-based activities gives safety a real chance to become part of day-to-day AI governance—not just policy theory.

🧩 What’s Missing

While the workbook is highly actionable, it’s narrowly scoped to public sector contexts and “narrow AI” systems. There’s little guidance on generative AI or foundation model safety, despite their growing use in government settings. Also, examples mostly focus on idealized risk scenarios. Adding content on navigating trade-offs (e.g., safety vs. speed, or robustness vs. transparency) would enhance its practical use. Lastly, some terms (like “model hardening” or “label flip rate”) could benefit from visualizations or simplified summaries for non-technical audiences.

✅ Best For

Ideal for civil servants, public sector tech leads, AI governance teams, and AI ethics trainers. Also useful for private sector orgs seeking to adapt public governance principles for internal use. Especially helpful if you’re setting up internal reviews, ethics boards, or integrating AI into legacy systems.