AI Ethics Governance: 9 Key Steps to Go from Zero to Operational
AI Ethics Governance: The Essentials in One Article — Real Code, Diagrams and Concrete Steps, Extracts from a 35-Lesson Course.
Everyone can learn AI Ethics Governance — provided they follow the steps in the right order. We have condensed a complete 35-lesson course into a clear pathway, with the most useful code excerpts.
- Introduction
- Fundamental Principles
- Algorithmic Biases
- Privacy and GDPR
- XAI Explainability
Why AI ethics now?
Learning objectives
Accelerator #1: the generalization of LLMs
In November 2022, the launch of ChatGPT marked a before and after. In two months, the tool reached 100 million users, an absolute record for a consumer service. Suddenly, every employee, every student, every journalist could generate text, code, and analyses.
The consequences are multiple:
Accelerator #2: the arrival of the EU AI Act
Adopted in March 2024, the Regulation (EU) 2024/1689 — nicknamed the “EU AI Act” — is the world’s first general AI law. It applies progressively between 2025 and 2027.
Overview
Timeline
Domino effect
The arrival of a binding standard changes everything: it is no longer about “doing one’s best,” but about documenting, auditing, and certifying.
Accelerator #3: societal and reputational pressure
The media, NGOs (Algorithm Watch, AI Now Institute, La Quadrature du Net), unions, and consumers scrutinize AI uses. A poorly managed incident becomes a scandal within hours on social media.
Recent examples:
Alert #1: COMPAS and the American justice system (2016)
The NGO ProPublica published an investigation in May 2016 on the COMPAS tool (Correctional Offender Management Profiling for Alternative Sanctions), used by American judges to estimate recidivism risk.
Alert #2: Amazon and sexist résumé screening (2018)
Reuters revealed that Amazon developed an internal résumé-screening tool between 2014 and 2017. The model, trained on 10 years of past applications (predominantly male in tech), systematically preferred men’s résumés.
The model penalized any mention containing the word “women’s” (women’s chess club, for example) and favored “masculine” verbs such as “executed” or “captured.” Amazon abandoned the project.
Alert #3: Tay and the fragility of chatbots (2016)
Microsoft launched Tay, a Twitter chatbot meant to learn “by chatting.” Within 24 hours, malicious users trained it to produce racist, Holocaust-denying, and sexist statements. Microsoft withdrew the tool.
The real cost of an AI failure
Fairness
Learning objectives
What is a “fair” AI?
The word fairness covers a complex reality: ensuring that a system does not unduly disadvantage certain people because of a protected attribute (sex, origin, age, disability, religion…).
Under French law, these attributes are listed in Article 225-1 of the Penal Code and the law of 27 May 2008. Any direct or indirect discrimination based on these criteria is prohibited, even when it originates from an algorithm.
The 4 major fairness measures
1. Demographic Parity
The positive prediction rate must be identical across protected groups.
Example: 30 % of men and 30 % of women are approved for a loan.
2. Equal Opportunity
Among people who deserve a positive outcome, the acceptance rate must be equal across groups.
Example: among qualified candidates, 80 % of men and 80 % of women are selected.
3. Equalized Odds
False positive and false negative rates must be identical across groups.
4. Calibration
When the model predicts a 70 % risk, 70 % of actual cases must be observed, identically for all groups.
Kleinberg’s impossibility (2017)
This result (Kleinberg, Mullainathan, Raghavan, 2017) explains why the COMPAS debate has no simple correct answer: Northpointe claimed the model was calibrated, ProPublica measured unequal error rates. Both were right… according to their own criterion.
How to choose a definition?
| Context | Recommended definition | Why |
|---|---|---|
| Recruitment | Equal Opportunity | Among the qualified, equal chance of being chosen |
| Credit | Calibration + monitored parity | Individual justice + group monitoring |
| Criminal justice | Equalized Odds | Presumption of innocence protected |
| Health (screening) | Sensitivity by subgroup | Do not miss cases |
| Targeted advertising | Demographic Parity (context-dependent) | Avoid discriminatory targeting (housing, employment) |
Concrete case: the 4/5 rule (US)
In the United States, the EEOC’s Uniform Guidelines on Employee Selection Procedures establish the “four-fifths rule” (1978): the selection rate for a protected group must be at least 80 % of the rate of the most selected group.
Famous cases — COMPAS, Amazon, Apple Card
Learning objectives
Case 1: COMPAS (Northpointe / Equivant)
The system
COMPAS — Correctional Offender Management Profiling for Alternative Sanctions — is a decision-support tool used since the early 2000s by American judges. It predicts recidivism risk based on 137 questions completed by the defendant and the probation service.
The score (1 to 10) is used to decide parole, bail amount, or sentence length.
The ProPublica investigation (2016)
ProPublica analyzed 7,000 scores in Florida and compared them with re-arrests over two years. Results:
Northpointe’s defense
Northpointe responded that its model is calibrated, hence fair for each individual. The debate reveals Kleinberg’s impossibility: calibration and equalized odds cannot both be satisfied when base rates differ.
The lessons
- Choose and document your fairness definition before deployment
- Score explainability must be published
- A proprietary (black-box) model in justice raises democratic issues
- Unequal base rates often reflect structural inequality
- Regularly audit with an independent third party
Case 2: Amazon recruitment (2014–2018)
The context
Amazon began developing an internal automated résumé-screening tool for tech positions in 2014. The model was trained on 10 years of applications, with the target “résumés resembling profiles hired over the past 10 years.”
The problem
The hired profiles were overwhelmingly male (the tech sector having historically been male-dominated). The model therefore learned to favor men.
More precisely:
The response
Amazon abandoned the project in 2017 and never deployed the tool “in production.” The information appeared in Reuters in 2018, becoming a global case study.
The lessons
- Never use past performance as a target when it reproduces inequality
- Removing sex is not enough: track proxies (vocabulary, degrees, clubs)
- Audit the model on subpopulations before any deployment
- Involve HR, legal, and ethics functions from the design stage
- Prefer decision-support (human in the loop) over 100 % automated screening
Case 3: Apple Card (2019)
The trigger
In November 2019, developer David Heinemeier Hansson tweeted that he and his wife share their accounts, yet she received a credit limit 20 times lower than his on the new Apple Card (issued by Goldman Sachs). The tweet went viral.
This article covers the most useful excerpts — the full AI Ethics Governance course (11 chapters, 35 lessons, corrected exercises and final project) takes you all the way.
./access-the-full-course free course: Claude CoworkFAQ
How long does it take to learn AI Ethics Governance?
Are there any prerequisites?
Where to start concretely?
📬 Want to receive this type of guide every week? Subscribe for free — real code, zero fluff.