The layer under the frameworks

The senior course "AI management system: ISO 42001" gave you the process: an inventory, a policy, an impact assessment, lifecycle controls, an evidence trail, a release gate, an audit. The senior course "Compliance: the EU AI Act" gave you the legal shape: risk tiers, provider and deployer roles, GPAI obligations and high-risk engineering work. Both courses can be followed to the letter. A team can complete every artifact either one requires: the inventory entry, the DPIA-style assessment, the conformity documentation. It can still ship a system that denies a loan to a group of applicants for no defensible reason. Or it explains a decision in language nobody affected can act on.

That gap is this course. It isn't a third framework to comply with. It's the substance the other two try to operationalize. ISO 42001 asks whether you have a process for assessing impact. It doesn't ask whether your fairness metric is the right one. Nor does it ask whether it even can be, given the other metrics you care about. The AI Act asks whether your high-risk system meets the accuracy and robustness obligations. It doesn't say what "fair" means when two reasonable mathematical definitions of fairness can't both hold for the same model. This course spends seven lessons on the part a checklist can't check.

What "ethics" means here

It isn't a values poster. It's four working questions a senior engineer has to answer with something more specific than "we care about this":

• Fairness: which of several incompatible mathematical definitions of "treated fairly" applies to this decision, and who decided that.
• Transparency: who needs to know what, at what level of detail, and which technique delivers that for them.
• Accountability: who answers for a harmful automated decision, and whether the person reviewing it can reverse it.
• The limits of process: what remains your judgement call even after every artifact above is filed correctly.

Why this isn't a repeat of what you already have

Here's a quick map, because the overlap is real but narrow:

If you haven't taken those two courses, you can still follow this one, because every idea here stands on its own. But this course won't re-teach what a management system is or how the AI Act's risk tiers work. It assumes you can build the paper trail. It teaches what should be true before you fill it in.

What this course covers

1. This lesson: where ethics sits relative to the compliance frameworks you already have, and the four working questions.
2. Fairness is not one thing: demographic parity, equalized odds and calibration are each a reasonable definition of "fair", and you generally can't satisfy all three at once. A worked example.
3. Where bias enters: historical training data, proxy variables, feedback loops, and evaluation sets that share the training set's blind spots.
4. Measuring it: disaggregated evaluation, why an aggregate number can improve while a subgroup gets worse, and a challenge that makes you compute it.
5. Three kinds of transparency: disclosure, explainability and interpretability are different obligations to different audiences, answered by different techniques.
6. Accountability, contestability and the record: who answers for a harmful decision, human review versus automation bias, and the record-keeping that makes an answer possible after the fact.
7. The limits of ethics as process: what a senior engineer is accountable for that no checklist encodes, and where you go next.

This course teaches engineering judgement, not legal advice. Nothing here tells you what your organization is legally required to do. That's counsel's job, working from the frameworks you already have. What follows is what a senior engineer should understand before that conversation, so the trade-off gets named instead of wished away.