# Exercises — Session 6 🟢 “AI and ethics”

**Program:** Applied AI — Beginner Level · **Instructor:** Yann Isola
**Materials:** a pencil, your phone (exercise 2), your critical thinking (all exercises)

---

## Exercise 1 — Bias detective 🕵️

**Duration:** 20 min · **Format:** alone or in pairs · **Difficulty:** ⭐⭐

Your mission: In each situation below, an AI produces a strange or unfair result. It's up to you to play detective: **find out where the bias comes from** and **propose a remedy**.

> 🔑 Reminder of the detective's magnifying glass: an AI learns from **data**. When she goes off the rails, the first question is always: "What was (or wasn't) in her training data?" »*

### Case #1: soap distracts
An automatic soap dispenser (with sensor) works great for some people...but doesn't detect the hands of dark-skinned people.

1. What is your hypothesis about the cause?
2. What would need to change to correct the problem?

### Case no. 2: the stereotypical translator
An automatic translator translates from a language where “doctor” and “nurse” have no gender. Systematic result: “**he** is a doctor”, “**she** is a nurse”.

1. Did the AI ​​“decide” this? Where does this reflex come from?
2. How could a translator present the translation more fairly? (Hint: some already do it — how?)

### Case no. 3: the prudent bank… too cautious
A credit AI rejects applications from a certain neighborhood more often — even when people have a good salary. The AI ​​never received the “neighborhood” information… but it knows the postal code.

1. How was the AI ​​able to “guess” the neighborhood when it was not given it directly?
2. Why would this case be classified as “high risk” by the AI ​​Act?
3. What should the person refused be able to do? (Think transparency…)

### Case No. 4: the image generator that lacks imagination
An AI is asked to generate “a photo of a CEO”. Out of 20 images, 19 show men in suits. We ask for “a person who does the housework”: almost only women.

1. Where does this imbalance come from?
2. Is it serious? Debate: “they’re just images” vs “images shape our ideas”.

### 🏆 Bonus question from the detective chief
Find something common to the 4 matters, and complete the golden rule:
*“An AI is never fairer than ________________. »*

### Corrected (do not read before searching!)

<details>
<summary>See the correction tracks</summary>

- **Case 1:** the sensor (and/or its calibration data) was mainly tested on light skin. Remedy: test and calibrate across the full diversity of real users, with diverse test teams.
- **Case 2:** the AI ​​learned from millions of human texts where “doctor” is more often associated with the masculine gender. It reproduces the statistic, not a truth. Remedy: offer **both translations** (“he/she is a doctor”) — some translators already do this.
- **Case 3:** the postal code is an **indirect index** (we say a “proxy variable”): it is correlated with the neighborhood, therefore sometimes with the social origin of the inhabitants. The AI ​​finds these shortcuts on its own. High risk because of a major financial decision over a lifetime.The person must be able to obtain an **explanation** and a **human review**.
- **Case 4:** training images (web photos, image banks) reflect decades of stereotypes. This is serious on a large scale: these images in turn feed the web… which the next AI will use to learn. Vicious circle possible.
- **Golden rule:** “…only the data that fed it. »

</details>

---

## Exercise 2 — Privacy audit of YOUR phone 🔍📱

**Duration:** 25 min (or at home) · **Format:** individual · **Difficulty:** ⭐⭐
**Addition:** the “Privacy Checker” tab of the session web page.

Today, you're the detective — and the suspect is your own phone. Objective: discover **what your apps know about you**, and regain some control. Without panicking: we notice, we understand, we adjust.

### Step 1 — Inventory (5 min)

List your **5 most used apps**:

| App | Free? | How do you think she makes money? |
|---|---|---|
| 1. | ☐ yes ☐ no | |
| 2. | ☐ yes ☐ no | |
| 3. | ☐ yes ☐ no | |
| 4. | ☐ yes ☐ no | |
| 5. | ☐ yes ☐ no | |

### Step 2 — The search for authorizations (10 min)

On your phone, open: **Settings → Privacy** (iPhone) or **Settings → Security & Privacy → Permissions Manager** (Android — the exact path varies by brand ⚠).

For each of your 5 apps, check what it has access to:

| App | 📍 Position | 📸 Camera | 🎤 Microphone | 👥 Contacts | 🖼️ Photos |
|---|---|---|---|---|---|
| 1. | ☐ | ☐ | ☐ | ☐ | ☐ |
| 2. | ☐ | ☐ | ☐ | ☐ | ☐ |
| 3. | ☐ | ☐ | ☐ | ☐ | ☐ |
| 4. | ☐ | ☐ | ☐ | ☐ | ☐ |
| 5. | ☐ | ☐ | ☐ | ☐ | ☐ |

### Step 3 — The interrogation (5 min)

For each box checked, ask THE detective's question:

> **“Does this app really NEED this access to do its job? »**

- A map app that asks for the position: logical ✅
- A puzzle game that requires your contacts and your microphone: 🤨 why?!

Write down your 2 most surprising discoveries:
1. _________________________________________
2. _______________________________________

### Step 4 — Regaining control (5 min)

Choose **at least 2 actions** and really do them:

- ☐ Remove an unnecessary authorization (you can always put it back afterwards!)
- ☐ Change a position from “Always” to “Only when the app is used”
- ☐ Delete an app that you no longer use (it may still be collecting)
- ☐ Look at the “privacy policy” of an app for 2 minutes and note something understood… or not understood at all
- ☐ (Advanced) Find out how to request a copy of your data from a major service — it's your right of access **GDPR**!

### Step 5 — The investigation report

Complete: *“Before this audit, I thought my apps knew ______. When I checked, I discovered ______. I decided to ______. »*

> 👨‍👩‍👧 **For younger children:** do this audit with a parent — you will probably be the expert on the settings, and he/she will be as surprised as you. The GDPR also provides enhanced protection for minors.

---

## Exercise 3 — The AI Tribunal ⚖️ (debate on ethical dilemma)

**Duration:** 20–30 min · **Format:** whole group · **Difficulty:** ⭐⭐⭐
**Addition:** the “Ethical Dilemmas” tab of the web page (voting and display of results).

### The case to be judged> **“The Jules-Verne college wants to use an AI to grade students’ papers and predict their chances of success. The court must decide: authorize, prohibit, or authorize subject to conditions. »**

### Roles

- ⚖️ **Judge(s)**: direct the debates, cut off time cheats, render a **motivated** verdict (the verdict must cite at least 2 arguments heard).
- 🟢 **The defense** (FOR the scoring AI): you are pleading for the college. Even if you think otherwise! Pleading an imposed position is muscle training for the brain.
- 🔴 **The accusation** (AGAINST): you plead for the association of worried parents.
- 👥 **The jury**: listens, asks questions, votes at the end (on the web page or by show of hands).

### The sequence (timed, it's important!)

1. Team preparation — **5 min**
2. Defense closing argument — **2 min**
3. Prosecution closing argument — **2 min**
4. Right of reply (1 min each) — **2 mins**
5. Questions from the jury — **3 min**
6. Jury vote + judges’ verdict — **3 min**
7. Debrief together — **3 min**

### Help sheet — Defense 🟢 (to cut out)

Tracks (choose, expand, add your own):
- An AI grades **all students with the same criteria** — no favoritism, no tired teacher on Sunday evening.
- **Faster** grading → teachers have more time to help students.
- Prediction allows you to **early spot** a student who drops out and help them BEFORE they fail.
- We can impose **guarantees**: a human reviews each note, the AI ​​is tested against bias.

### Help sheet — Accusation 🔴 (to cut out)

Tracks:
- The AI learns from **old notes**… which may have already contained injustices. The bias is copied, then amplified.
- **Black box**: how can you challenge your rating if no one can explain why?
- Student data is **data of minors** → maximum protection (GDPR!).
- A **prediction** can become a label: “the AI ​​said you will fail”… what if we ended up believing it?
- The European AI Act classifies AI in education as **high risk** — this is no coincidence.

### Questions the jury can ask (help)

- “Who will be responsible if the AI makes a mistake on a note in the patent? »
- “Have the parents given their consent? And the students? »
- “What happens to a student with an original writing style that the AI has never seen? »
- “What specific guarantees do you offer? »

### The debrief (the most important part!)

1. Those who argued against their opinion: how did you feel? Did you find any good arguments anyway?
2. Did anyone **change their mind** during the trial? By what argument?
3. Is there a **median** solution (allow under conditions)? Which ?
4. 🎬 **Professor's revelation:** this affair has already happened in real life (notes by algorithm in the United Kingdom in 2020 → injustices → demonstrations → abandonment). What does that change to your verdict?

### Additional cases (if the court wants to sit again)

- 📹 “Should facial recognition cameras be installed at the college entrance for security? »
- 🏥 “Should a village without a doctor equip itself with diagnostic AI that can be used without a doctor? »
- 💬 “Should we allow “virtual friends” chatbots for people who feel alone? »
- 🌍 “Should we limit the use of large generative AI for ecological reasons? »> ⚖️ **Moral of the court:** on these questions, there is not ONE good answer hidden somewhere. There are arguments, values, possible guarantees — and citizens (you!) who will have to decide. This is exactly why we are debating it.