Skip to main content

Ethics and responsible AI

How to move from principles to verifiable requirements, and audit a real machine learning system end to end.

Course Duration: 4h​

What You'll Learn​

  • Turn abstract principles (fairness, transparency, accountability, privacy) into requirements you can measure
  • Locate the origin of bias in the data pipeline and quantify its impact
  • Choose fairness metrics deliberately and understand why they cannot all hold at once
  • Explain a model's decisions with SHAP and LIME, and know what those explanations do not prove
  • Protect personal data with anonymization, k-anonymity and differential privacy
  • Write a model card and a datasheet that a compliance officer will actually read
  • Classify a system under the EU AI Act and list the obligations that follow
  • Run a complete audit and produce a corrective action plan

Prerequisites​

  • Supervised learning (course 04): training, evaluation, common metrics
  • Reading Python code that manipulates tabular data with pandas

The through-line: a consumer credit scoring model​

Every module returns to the same case: a bank scores retail credit applications with a gradient boosting model. Module 1 turns the bank's principles into concrete requirements. Module 2 tracks the bias that decades of past decisions injected into the training set. Module 3 measures fairness with Fairlearn and shows why demographic parity and equalized odds cannot both be satisfied at once. Module 4 uses SHAP to open the box, then explains why an explanation is not a justification. Module 5 shows how ZIP code plus date of birth re-identifies almost anyone. Module 6 turns the audit into a model card and a datasheet. Module 7 designs the recourse procedure a rejected applicant can invoke. Module 8 places the system under the EU AI Act (high-risk, since it scores creditworthiness) and enumerates the obligations. Module 9 covers personal data protection under the GDPR, including the article 22 rules on automated decisions. Module 10 stitches everything into an audit report.

Course Modules​

  1. What "responsible" concretely means
  2. Sources of bias: collection, labeling, history
  3. Fairness metrics and their incompatibilities
  4. Explainability: LIME, SHAP and their limits
  5. Privacy: anonymization and differential privacy
  6. Documentation: model cards and datasheets
  7. Human oversight and recourse
  8. The EU AI Act and risk levels
  9. Personal data protection
  10. Case study: auditing a decision system

Assessment and certificate​

The course ends with a 40-question exam covering every module. On success, a certificate of completion is issued; its number is verifiable on the platform.

Free courses, by contrast, end with a 5-question quiz and a preview of the certificate, without certification.