AI curriculum

Course outline for Responsible AI

An AI recommendation can harm people who never chose to use the system. Review a fictional feature for unequal effects, unsupported claims, data exposure, and meaningful human choice. Define a reachable challenge process and a bounded deployment decision that leaves unresolved harms visible.

Technology: AI Security, Governance and Trust
Duration: 6 hours
Modules: 3 · 2 hours each
Level: Level 2

About Responsible AI

Responsible AI Course Objectives

Identify affected stakeholders beyond the immediate user.
Compare fairness and bias concerns using supplied outcome examples.
Distinguish transparent evidence from persuasive unsupported explanation.
Define privacy, safety, human-choice, and fallback safeguards.
Route an adverse result through challenge, correction, and deployment review.

Pre-requisites

AI capability and limitation awareness equivalent to Artificial Intelligence Foundations.
Interest in organizational or product decisions and effects on people.
Coding is not required; workflow modeling is helpful for mapping review and fallback ownership.

Before this course

Complete the required courses or bring equivalent experience. Recommended courses add useful depth; they are not enrolment gates.

Required preparation

Artificial Intelligence Foundations

Recommended preparation

Workflow Foundations for AI Automation

Understands AI limitations before analyzing effects on people, privacy, and meaningful human choice; coding is not required.

Find your place in the learning map

Lab Setup

Computer with a document editor and optional spreadsheet.
Use supplied fictional scenarios, synthetic outcome tables, and guided model-output comparisons.
No coding, personal data, model account, or live consequential decisions are required.

Detailed Course Outline

Module 1: People and uneven effects

Ethics: Map affected stakeholders, competing interests, and consequential harms; include nonusers and record unresolved value judgments in the deployment decision.
Fairness: Compare unequal outcomes across supplied stakeholder groups; identify competing fairness concerns and state what a small synthetic sample cannot establish.
Bias: Inspect representation and outcome patterns in supplied examples; distinguish observed imbalance from established causes and avoid unsupported population claims.

Practical outcome: Produce a stakeholder/harms map from synthetic outcomes, identify representation and outcome concerns, and state what the small sample cannot establish.

Module 2: Evidence and meaningful choice

Transparency: Compare source-supported recommendations with persuasive unsupported explanations; specify disclosures about evidence, uncertainty, and limits that support informed decisions.
Safety: Identify consequential failure paths and appropriate safeguards; require an owner, fallback, or explicit unresolved-risk decision for each material harm.
Privacy: Map data collection, sharing, retention, and exposure in the fictional feature; distinguish local processing from an actual privacy assessment.

Practical outcome: Inspect a plausible unsupported recommendation and specify evidence, data-sharing, and safety safeguards; distinguish local processing from an actual privacy assessment.

Module 3: Challenge and bounded deployment

Human agency: Provide meaningful alternatives, reachable reviewers, and correction paths; distinguish genuine choice from passive acceptance of an automated recommendation.
Responsible deployment: Set bounded release conditions, accountable owners, and stop triggers; retain unresolved harms rather than converting incomplete evidence into approval.
Contestability and harm feedback: Route an adverse outcome through challenge, review, correction, and deployment reconsideration; preserve feedback and name owners for manual fallback.

Practical outcome: Define reachable reviewers, alternatives, correction ownership, retained expertise, and harm-feedback triggers for a conditional release or stop decision.

Practical exercise

Submit a deployment review, stakeholder map, and challenge/fallback walkthrough for a fictional recommendation feature.
Review one supported result, one unsupported but persuasive result, and one adverse outcome affecting a nonuser stakeholder.
Give every identified consequential harm a safeguard or an explicit unresolved-risk decision and accountable owner.
Reject the unsupported recommendation, route the adverse case to a reachable reviewer, and document correction or manual fallback before the release decision.

What you will produce

Design a responsible deployment review with actionable challenge and fallback paths.

How the work is checked

The stakeholder map includes affected people beyond the user and states sample limitations.
The unsupported result is rejected and every consequential harm has an owner and disposition.
The adverse-case walkthrough reaches a reviewer and a correction or fallback decision.

Six-hour noncoding review; quantitative fairness evaluation, ethical frameworks, and live safety assurance require additional preparation.

When another approach is better

Passive acceptance is not meaningful oversight when users cannot challenge a result or seek correction.
One fairness measure or a small synthetic sample cannot settle all stakeholder concerns.
Do not deploy a consequential feature when serious unresolved harms lack a responsible owner or fallback.

How we train

Ask about exercises, instructor feedback and the prior knowledge you need. Tool-specific courses marked provisional may become modules in a broader course.

Discuss your learning goals