Recap and final exam
Ten modules to move from a corporate charter to a signed audit report on a real decision system. Here is the course condensed, then the threads that run through it.
The course at a glance
| Module | The essential point |
|---|---|
| 1. What responsible means | Principles are worthless until they become verifiable requirements: a metric, a threshold, a document, a named owner |
| 2. Sources of bias | Bias enters at collection, measurement, labeling, feature engineering and deployment; removing the protected attribute leaves the proxies |
| 3. Fairness metrics | Demographic parity, equalized odds and calibration cannot all hold when base rates differ (impossibility theorem) |
| 4. Explainability | SHAP is stable and additive, LIME is unstable; an explanation is not a justification |
| 5. Privacy | Quasi-identifiers re-identify most people; differential privacy replaces removal with a bounded budget |
| 6. Model cards and datasheets | The intended-use and out-of-scope sections prevent scope creep and shift accountability when misuse happens |
| 7. Oversight and recourse | Human in the loop is a rubber stamp without countermeasures against automation bias; recourse means deadlines, independent reviewers and counterfactuals |
| 8. EU AI Act | Credit scoring is high-risk (Annex III 5(b)); seven obligations apply from 2 August 2026, with extraterritorial reach |
| 9. GDPR | Article 22 forbids solely automated significant decisions without safeguards; a DPIA is required and must be redone on substantial change |
| 10. Case study | The audit grid produces a Yes/No/Partial verdict per section; blockers stop production sign-off |
The threads running through the course
Bias is a data problem, not an algorithm problem. Modules 2, 3 and 4 keep returning to the same claim: the model reproduces and amplifies what the data contains. That is why the audit spends more time on data provenance and label semantics than on model architecture, and why the fairness fix is rarely "try a different classifier".
Every principle becomes an artefact. Fairness is a Fairlearn output with a threshold. Transparency is a SHAP waterfall in a rejection letter. Accountability is a name on the model card. Privacy is a k-anonymity level plus a differential-privacy budget. Human oversight is a HITL/HOTL split with a disagreement dashboard. Compliance is a DPIA and an Annex IV dossier. If the principle has no artefact, it is a slogan.
Regulation is not optional and not new. The AI Act and the GDPR do not invent responsibility — they codify it and add penalties. Modules 8 and 9 spend most of their time on obligations that a well-run project would satisfy anyway; the two regulations mostly ensure that badly-run projects cannot ignore them.
An audit is a stop-and-check, not a badge. Module 10 makes the point explicit: the credit scorer will be audited again in eighteen months, and again after that. Responsible AI is a periodic public review, not a one-time certification.
The final exam
The exam has 40 questions covering the ten modules: turning principles into verifiable requirements; identifying selection, measurement, label and proxy bias in concrete scenarios; choosing between demographic parity, equalized odds and calibration under the impossibility constraint; reading SHAP and LIME outputs and spotting instability; sizing k-anonymity and reasoning about differential-privacy budgets; writing model card sections and datasheet fields; designing HITL/HOTL splits and recourse procedures; classifying systems under the AI Act and enumerating their obligations; applying GDPR principles including Article 22 to automated decisions; running the audit grid on a fresh case.
Several questions present situations to diagnose: a "cleaned" dataset that still leaks the protected attribute, a fairness dashboard that hides a subgroup gap, a rejection letter that violates Article 22, an audit dossier missing a section. It is judgment that is assessed, not the recitation of definitions.
On success, your certificate of completion is issued immediately; its number is verifiable by any third party on the platform.
Take the table above and, for each row, ask yourself "how would I see that I am wrong here?". If you can say why removing gender does not remove discrimination, why equalized odds and calibration cannot both hold, and what changes when a solely-automated decision becomes solely a human decision, you are ready. Good luck!
Final exam
Ready to validate this course?
40 questions drawn at random from the course bank · passing score 70% · verifiable PDF certificate issued immediately on success.
Start the examYou need to be signed in to your InSkillML account with an active subscription. You can also start the exam from My courses.