Skip to main content

Module 1 — What "responsible" concretely means

Every large organization now publishes a page of AI principles. They read almost identically: fairness, transparency, accountability, safety, privacy. Read one, you have read them all. The problem is that a principle you cannot measure protects no one. A model deployed under a fairness principle will still discriminate if nobody defined which metric fairness would be evaluated against, on which subgroups, above which threshold, and who is responsible for pulling the plug when the threshold is crossed.

This course is about that translation work. From this module on, every principle appears as a verifiable requirement: a metric with a threshold, a document with a template, or an action with a designated owner. If the requirement cannot be checked, it does not enter the model card.

The through-line: a credit scoring model at a mid-sized bank​

The bank we will follow trains a gradient boosting model on twelve years of retail credit applications. The label is whether the loan defaulted within twenty-four months. The features are income, employment, existing debt, ZIP code, age, and a handful of derived variables. The model outputs a probability of default; an application is approved when the probability is below a threshold. The bank's board has committed to a "responsible AI" charter. Our job for the next nine modules is to turn that charter into an audit report that either clears the model for production or lists what must change first.

Five principles, and what they demand in practice​

Fairness requires that error rates and approval rates do not systematically disadvantage a protected group. In practice, that becomes: pick a fairness metric (module 3 will show it cannot be all of them at once), define the subgroups (gender, age band, ZIP-code proxy for neighborhood), set an acceptable disparity threshold, and measure it on a held-out set that mirrors the applicant population. If the threshold is exceeded, the model does not ship.

Transparency requires that a rejected applicant understand why. That is not a slogan — it is a document. For each rejection, the system must produce the top contributing features (module 4 uses SHAP for this), in language the applicant can read, together with what would have to change for the decision to flip. That document is legally required in some jurisdictions and morally required everywhere.

Accountability requires that a named human owns the decision. Not "the model decided" — the risk committee decided to deploy this version of the model, on this segment, with this threshold, on this date. When something goes wrong, an org chart entry, not a URL, answers for it.

Safety requires that the system fail in predictable ways. What does the API return when a feature is missing? When the model is unavailable? When a request contains an obviously adversarial input? A "responsible" system has an answer to each; a merely trained model does not.

Privacy requires that the data used to train, evaluate and monitor the model does not leak the identity or the sensitive attributes of the people in the training set. Modules 5 and 9 cover the two halves of this: the statistical side (anonymization, differential privacy) and the legal side (GDPR).

Stakeholders — and the ones easy to forget​

A responsible-AI audit lists the parties whose interests the system affects. The visible ones are the users (the loan officers), the customers (the applicants), the deploying organization (the bank), and the regulator. The ones consistently forgotten are the rejected applicants, the third parties whose data appears in shared households, and the labelers whose decisions shaped the historical training data. When we get to module 2, we will see that ignoring the last group is how bias enters undetected.

What this course does not promise​

Responsibility is not a technical property that can be certified once. There is no library import that makes a model ethical, and no metric threshold that guarantees fairness across contexts the model was never tested on. What the course delivers is a workflow: a chain of checks and documents that, followed rigorously, makes an unethical deployment much harder — and, when one happens anyway, makes the failure legible so that it can be corrected and explained.

The course also does not decide the political questions. Whether a bank should use algorithmic scoring at all, or whether a particular subgroup disparity is "acceptable", are decisions for humans with democratic legitimacy, not for the model or its authors. The audit informs those decisions; it does not replace them.

Summary​

  • Principles are worthless until they become verifiable requirements: a metric, a threshold, a document, or a named owner.
  • The course's through-line is a credit scoring model to be audited end to end across modules 2 through 10.
  • Fairness, transparency, accountability, safety, privacy each translate into concrete artefacts: subgroup metrics, per-decision explanations, org-chart ownership, failure-mode specs, privacy budgets.
  • Rejected applicants, third parties and labelers are stakeholders too — omitting them is how bias becomes invisible.

Next module: where bias comes from in the credit dataset, and why the model can be unbiased in code and biased in effect.