AI Ethics in Practice: Fairness, Transparency and Accountability
Find where bias enters AI systems, measure fairness in Python, explain and document decisions, and set up the governance and audits that keep AI accountable.
Intermediate · Responsible AI · about 11 hours · 4 modules · certificate on completion
Taught by Amahlé Ndlovù, Deshawné Thompsòn, AI Learning Companions.
About this course
Ethical AI is decided by concrete choices: what data to use, which errors to accept, who gets an explanation and who answers when things go wrong. This course turns principles into practice. You'll map the people an AI system affects and the harms it could do, trace how bias enters through historical data, under-representation, proxies and deployment, and measure fairness in pandas with selection rates, the disparate impact ratio and equalised odds. You'll see why fairness measures can't all be met at once when groups differ, and how to make and record a defensible choice. You'll explain individual decisions with reasons and counterfactuals, document a model with a model card, and apply privacy principles to AI projects. Finally, you'll set up accountability and governance, from human oversight to impact assessments, and audit a lending model's decisions by group, by threshold and at the intersection of two characteristics. The labs use a small, made-up set of loan decisions, and the course ends with a full audit report.
What you'll learn
- Apply widely shared ethical principles for AI to concrete decisions about building and using AI systems
- Trace where bias enters an AI system, and measure it with demographic parity, equalised odds and related measures in pandas
- Explain why fairness measures can conflict, and make and document a reasoned choice between them
- Explain individual decisions to the people affected, and document models with model cards
- Set up accountability and governance processes, and run a fairness audit from scope to report
Skills you'll gain
AI ethics, Bias analysis, Fairness metrics, Explainability, Model documentation, Privacy by design, AI governance, Algorithmic auditing
Tools
Python 3, pandas
Before you start
Basic Python and pandas: reading a CSV file, selecting rows and using groupby. Install Python 3 and pandas (pip install pandas). No background in ethics or law is needed.
Modules
Ethical principles for AI
Why AI ethics is a practical matter, the principles that frameworks around the world share, and how to turn them into decisions about a real system.
- Why AI ethics is a practical matter (10 min)
- Principles for trustworthy AI (12 min)
- From principles to decisions (10 min)
- Hands-on lab: Map the stakeholders and harms of a lending model
Where bias comes from, and measuring fairness
Trace how bias enters a system, measure it with selection rates, the disparate impact ratio and equalised odds, and see why fairness measures can't all be satisfied at once.
- Where bias comes from (12 min)
- Measuring fairness: demographic parity and equalised odds (14 min)
- Why fairness measures conflict (12 min)
- Hands-on lab: Measure fairness in loan decisions with pandas
Transparency, explainability and privacy
Document models honestly, explain individual decisions with reasons and counterfactuals, and apply data-protection principles and privacy techniques to AI projects.
- Transparency and documentation (10 min)
- Explaining model decisions (12 min)
- Privacy in AI systems (12 min)
- Hands-on lab: Explain a decision and write a model card
Accountability, governance and auditing
Make named people answerable for AI systems, build governance processes across the life cycle, and audit a model's fairness from scope to findings.
- Accountability and human oversight (10 min)
- Governance processes (12 min)
- Running a fairness audit (12 min)
- Hands-on lab: Audit thresholds and intersections
Project: A fairness audit of a lending model
The lender's board has asked for an independent-style audit of its loan scorecard before it decides whether to keep using it. Using loans.csv and the scorecard from the labs, audit the model's decisions and write a report for the board. Set the scope and criteria first; measure the model by group, by age band and at their intersections; test at least two threshold options and explain their trade-offs; review its transparency (with a model card and a sample explanation for a declined applicant); and cover privacy, accountability and governance: how protected-attribute data is handled, who oversees decisions, how applicants can challenge them and what is monitored. End with a recommendation to keep, change or stop the model, stating the trade-offs openly. The data is made up, so be clear about what a real audit would need that this data can't show.
Part of
The products this course names are trademarks of their owners, who haven't made or endorsed it.