Module 24

Responsible AI

leakage, bias, explainability, privacy, prompt injection, model card

When you finish this module you can say:

Before I trust a model I can check it for leakage, for gaps between groups, for what it looks at, for privacy and for language model security, and write what I found on a model card.

By the end of this module you will have an audit function that catches leakage and group gaps in a model's measurement, code that cleans personal data and defends in layers against prompt injection, and a model card that explains what you found.

Lessons

6 lessons · ≈ 72 min

1. Data leakage and honest measurement12 min
2. Measuring bias12 min
3. Explainability: what does the model look at?12 min
4. Privacy: cleaning personal data12 min
5. Language model safety: prompt injection12 min
6. Project: model card and automatic audit12 min · project

Practice exercises

Short writing exercises, independent of the lessons: Go to the exercises

1. Find the suspicious columns
2. Selection rate by group
3. Mask the ID number