Responsible AI, Ethics & Governance

What you’ll be able to do

1. LO1Identify the stakeholders and harms (allocative, quality-of-service, representational, privacy, safety) of an AI system, critique its target variable, and run a risk and impact assessment that ends in a justified build, change or stop decision.
2. LO2Trace how bias enters across the ML lifecycle (historical, representation, measurement, aggregation, evaluation, deployment, feedback loops), detect proxies quantitatively, and document data with datasheets and data statements.
3. LO3Compute, interpret and choose group and individual fairness metrics with confidence intervals, and derive why independence, separation and sufficiency cannot all hold when base rates differ.
4. LO4Implement and compare pre-, in- and post-processing mitigations and evaluate their fairness-accuracy trade-offs on held-out data, including levelling down and delayed impact.
5. LO5Implement permutation importance, partial dependence and ICE, exact and sampled Shapley values, and actionable counterfactual explanations, and state when an explanation is unfaithful or misleading.
6. LO6Demonstrate linkage and membership-inference attacks, and apply k-anonymity and differential privacy (randomised response, Laplace and Gaussian mechanisms, DP gradient descent) with correct privacy accounting.
7. LO7Detect distribution shift, mount and defend against adversarial examples with honest evaluation, and analyse poisoning, supply-chain and generative-AI misuse risks, including detector base rates and watermark tests.
8. LO8Map an AI system's obligations under India's DPDP Act 2023 and Rules 2025, the IT Rules, national AI guidelines and sector regulators (RBI, CERT-In), and the copyright and licence terms of its training data, and compare them with the EU AI Act and other regimes, stating commencement dates with a primary source.
9. LO9Produce audit evidence and accountability artefacts (model cards whose numbers recompute, audit reports, release gates, incident-response plans) that a reviewer can check independently.

Chapters

6 chapters · each section ends with a short check

1. 01Harms, targets and where bias enters
2. 02Measuring and mitigating unfairness
3. 03Explaining models and the limits of explanation
4. 04Privacy: anonymisation failures, differential privacy and attacks on models
5. 05Robustness, security and generative-AI risks
6. 06Governing AI: India's laws and institutions, the EU AI Act, copyright, audits and incident response

Graded labs

1. L1

   Data, targets and proxies

   After chapter 1 · about 5 hours · Synthetic loan applications · Course own work, CC BY 4.0 · Runs on a CPU

2. L2

   Fairness audit and mitigation

   After chapter 2 · about 6 hours · Synthetic loan applications · Own work, CC BY 4.0 · Runs on a CPU

3. L3

   Explanations from scratch: importance, Shapley values and recourse

   After chapter 3 · about 5 hours · Diabetes · BSD-3-Clause (scikit-learn) · Runs on a CPU

4. L4

   Breaking and protecting privacy

   After chapter 4 · about 5 hours · Synthetic district-hospital discharge release · Course own work, CC BY 4.0 · Runs on a CPU

5. L5

   Release review for a PIN-code digit reader: robustness, shift and incident readiness

   After chapter 6 · about 5 hours · Optical recognition of handwritten digits · BSD-3-Clause (scikit-learn) · Runs on a CPU

Before you start

Where it leads

How it’s taught

You read short sections written for the page. Each ends with a check, and in-browser exercises let you practise. Graded labs run on your own cloud lab machine, CPU-friendly.

No videos. No live classes. No AI tutor.

How it works→