AI Ethics, Risk & Compliance

1. 01

   Responsible AI foundations

   Beginner

   The vocabulary and the recurring failure patterns — established before the frameworks, and the reason the frameworks exist.

   • Fairness and Machine Learning Book fairmlbook.org Barocas, Hardt, Narayanan 15h Free Free in full and the standard academic text. Chapters 1 and 2 are readable without a maths background and worth everyone's time.
   • Google — Responsible AI practices Documentation Google 3h Free A practitioner's checklist rather than a philosophy. Useful for what to look at; pair it with the academic sources for why.
   • Gender Shades Paper MIT Media Lab Joy Buolamwini, Timnit Gebru 1h Free The study that made disparate performance across demographic groups impossible to dismiss. Read the method, not just the finding.
2. 02

   Bias, fairness and evaluation

   Advanced

   Measuring disparate impact rather than asserting there is none — including the result nobody likes: several reasonable fairness definitions are mathematically incompatible.

   • Fairlearn — User Guide Documentation Fairlearn 5h Free Open-source fairness assessment with worked notebooks. The place to move from reading about metrics to computing them.
   • Aequitas — Bias and Fairness Audit Toolkit Code repository Carnegie Mellon University 3h Free Built for public-policy contexts, where the consequences of getting this wrong are most direct. The fairness tree is an excellent decision aid.
   • Datasheets for Datasets Paper arXiv Gebru et al. 1h Free The questions you should be able to answer about any dataset you use. Most teams cannot answer half of them, which is the point.
   • Model Cards for Model Reporting Paper arXiv Mitchell et al. 1h Free The companion standard for the model itself: intended use, out-of-scope use, and performance broken out by group.

   Build this

   A bias audit with numbers in it

   Take a public model or one you have trained. Split performance by a protected attribute, compute several fairness metrics, and show where they disagree with each other.

   Then write the hard part: which definition you would choose for this application and why the others are wrong for it.

   Done when: An audit report with disaggregated metrics and a defended choice of fairness definition.

   UCI Adult, COMPAS, or any dataset with documented protected attributes
3. 03

   Privacy and data governance

   Intermediate

   What you are allowed to hold, for how long, and what a model trained on it can leak back out.

   • General Data Protection Regulation — full text Documentation gdpr-info.eu 5h Free Articles 5, 6, 22 and 35 are the ones that bear on AI: principles, lawful basis, automated decision-making, and impact assessments.
   • Programming Differential Privacy Book programming-dp.com Joseph Near, Chiké Abuah 8h Free Free, code-first, and it makes a genuinely difficult topic tractable. Read it before claiming any dataset is anonymised.
   • ICO — Guidance on AI and data protection Documentation Information Commissioner's Office 5h Free A regulator explaining, concretely, what compliance looks like in an AI system. Far more useful than most commentary about it.
4. 04

   Security of AI systems

   Advanced

   The attack surface that only exists because the system takes instructions in natural language.

   • OWASP Top 10 for LLM Applications Documentation OWASP 3h Free Prompt injection, insecure output handling, excessive agency. The reference list every LLM application review should start from.
   • MITRE ATLAS Documentation MITRE 4h Free Adversarial tactics against ML systems, mapped in the same structure as ATT&CK, with real case studies attached.
   • NIST — Adversarial Machine Learning taxonomy Paper NIST 4h Free The shared vocabulary for attack types. Dense, and the reference you will keep returning to when writing threat models.

   Build this

   A threat model for an LLM application

   Take a real assistant or agent — ideally one your organisation uses. Work through the OWASP list and, for each item, write down whether it applies, what the impact would be, and what control exists today.

   Where no control exists, say so plainly. An honest gap list is the deliverable; a reassuring one is worthless.

   Done when: A threat model with an honest, prioritised list of unmitigated gaps.

5. 05

   Governance frameworks and standards

   Advanced

   The frameworks your risk, legal and audit functions will actually invoke — and how to document a system so it passes.

   • NIST AI Risk Management Framework Documentation NIST 5h Free Read the framework and the Playbook together. Govern, Map, Measure, Manage is the structure most enterprise programmes now use.
   • EU AI Act Explorer Documentation EU AI Act Explorer 4h Free Browsable article by article with recitals attached. Know the risk tiers and which obligations attach to each.
   • ISO/IEC 42001 — AI management systems Documentation ISO 2h Free The certifiable management-system standard. The full text is paid; the overview tells you what an audit against it would ask for.
   • OECD AI Principles Documentation OECD 2h Free The intergovernmental baseline most national AI strategies, India's included, are written in reference to.

   Build this

   A documentation pack for one AI system

   Produce, for a single real system: a model card, a datasheet for its training or retrieval data, and a risk register mapped to the NIST framework's four functions.

   This is precisely the pack an enterprise procurement or audit process asks for, and being able to produce one is a hireable skill on its own.

   Done when: A model card, a datasheet, and a mapped risk register for one system.

6. 06

   Writing policy people follow

   Intermediate

   The last step, and the one that fails most often: rules specific enough to apply and permissive enough to be obeyed.

   • GOV.UK — Writing for the web Documentation GOV.UK 2h Free How to write rules that people under time pressure can follow correctly. Policy fails as writing far more often than as policy.
   • NIST AI RMF Playbook Documentation NIST 4h Free Suggested actions against each framework outcome. The nearest thing to a template for the controls a policy needs to name.

   Build this

   An AI usage policy an organisation would adopt

   Two pages, maximum. What is permitted, what requires review, what is prohibited, what must be disclosed, and who decides the edge cases.

   Then test it: give three colleagues a genuinely ambiguous scenario and see whether they reach the same answer from your text. If they do not, the policy is not finished.

   Done when: A two-page policy, plus evidence that three readers applied it consistently.

Nothing matches that combination in this path yet.