# Teacher's Guide — Session 6 🟢 # “AI and ethics” **Program:** Applied AI — Beginner Level (from 12 years old, general public) **Instructor:** Yann Isola **Recommended duration:** 2 hours (adaptable 1h30 – 2h30) **Prerequisites:** Sessions 1 to 5 (concepts: AI, machine learning, data, generative AI, deepfakes) --- ## 1. Educational objectives At the end of the session, each participant should be able to: 1. **Explain** in your own words why an AI can be biased (“AI learns from data; if the data is biased, the AI will be biased”). 2. **Cite** at least 3 concrete examples of bias in AI (facial recognition, recruitment, targeted advertising). 3. **Describe** where the data that trains AIs comes from — often from us! — and name 4 types of personal data collected. 4. **Explain** what consent is and what the **GDPR** (General Data Protection Regulation) guarantees in Europe. 5. **Define** the “black box” problem and explain why transparency is crucial (health, justice, banking). 6. **Discuss** the question of responsibility (e.g.: autonomous car) by distinguishing between the different possible actors. 7. **Locate** the environmental impact of AI (energy, water, CO2) with orders of magnitude. ⚠ 8. **Nuance** the AI/work relationship: transformed jobs, threatened jobs, created jobs — AI as a tool, not as a replacement. 9. **Quote** the European **AI Act** ⚠ and its 4 levels of risk (unacceptable, high, limited, minimal). 10. **Participate** in a structured debate (“AI court”) by arguing a position, even if it is not your own. > ⚠ **Volatile content**: energy consumption figures, the state of application of the AI ​​Act, and current examples are evolving. Check before each session (15 min of wakefulness). The sections marked ⚠ should be re-checked as a priority. --- ## 2. Overview and common thread **Code of the session:** *“AI is neither nice nor bad: it is a mirror. It reflects our data, our choices… and our faults. Today, we learn to look in this mirror with a critical eye. »* This is THE citizen session of the program. No new technique: we mobilize everything that has been seen (data, learning, generation) to ask the question **“Is this good?” is this right? who decides? »**. The highlight is the **AI Court**: a dramatized debate where participants argue for or against the use of AI. **Teacher's posture — golden rule:** on ethical dilemmas, **you never impose THE right answer**. You animate, you relaunch, you demand arguments. Many of the questions in this session have no single answer — and that's exactly how participants should feel. ### Timed course (based on 2 hours)| Time | Sequence | Format | |---|---|---| | 0:00 – 0:10 | Tagline: “Can AI be unfair? » | Discussion | | 0:10 – 0:30 | Biases: where do they come from? Real-world examples | Presentation + **bias simulator** (web page) | | 0:30 – 0:45 | Privacy & consent: our data, the GDPR | Presentation + **privacy checker** (web page) | | 0:45 – 0:55 | The black box: transparency and trust | Presentation + discussion | | 0:55 – 1:05 | **Pause** | — | | 1:05 – 1:15 | Responsibility: the case of the autonomous car | Lightning mini-debate | | 1:15 – 1:25 | Environment + work: the two impacts | Presentation ⚠ | | 1:25 – 1:35 | The laws: GDPR + AI Act, the 4 levels of risk ⚠ | Exposed | | 1:35 – 1:55 | **Activity: AI Tribunal** (exercise 3) | Structured debate + **dilemma cards** (web page) | | 1:55 – 2:00 | Summary + quiz (or quiz at home) | Quiz | --- ## 3. Detailed content, sequence by sequence ### 3.1 Hook (10 min) — “Can AI be unfair? » Launch with a true story, told simply: > *“A large company built an AI to sort CVs and find the best candidates. After a few months, we discovered a problem: the AI ​​was almost systematically excluding women's CVs. No one told him to do this. So… where did the problem come from? »* Let it look. Expected answers (to emerge): - “She learned from old CVs” ✅ — that’s exactly it - “Someone programmed her badly” ❌ — no, no one coded “rejects women” - “She’s stupid” — partially: she doesn’t *understand*, she imitates **Reveal the key to the session:** the company had recruited mostly men for 10 years. The AI, trained on this historical data, deduced that “good candidate = male profile”. She even learned to penalize words like “women’s team…”. > 💡 **Phrase to remember from the session:** “An AI is never more accurate than the data that fed it. » **Pedagogical trap to avoid:** do not transform the session into “AI is evil”. The message is: *AI is a powerful tool; like any powerful tool, you need rules, vigilance and citizens who understand.* --- ### 3.2 Biases in AI (20 min) #### The mechanism (to be explained with the analogy of the cook) > *“Imagine an apprentice chef who learns only by watching a single chef. If this chef salts his dishes too much, the apprentice will salt too much — without ever wondering if it is good. AI is this apprentice: it reproduces what we show it, including the faults. »* Logical chain to write on the board: **Imperfect world → imperfect data → AI that learns these imperfections → imperfect decisions… at scale and very quickly.** The bottom line: bias is not a “failure.” The AI ​​works *perfectly* — it faithfully reproduces what it saw. That's what's tricky. #### Three real examples to tell 1. **Facial recognition**: studies (notably that of researcher Joy Buolamwini at MIT) have shown that certain facial recognition systems make mistakes much more often on the faces of dark-skinned women than on the faces of light-skinned men — sometimes less than 1% error for some, more than 30% for others. For what ? The training photos contained mostly faces of light-skinned men.Real-world consequences: Innocent people have been arrested due to misidentification. 2. **Recruitment**: the example of the catchphrase (the CV sorter which disadvantaged women). The company ended up abandoning the tool. 3. **Targeted advertising**: studies have shown that ads for well-paid jobs (managers, engineers) are shown more often to men than to women; or that advertisements for expensive credits were more targeted at certain neighborhoods. The algorithm reproduces and **amplifies** existing stereotypes. #### Demo: the bias simulator (web page, tab 1) Open `webpage/index.html`, tab **“Bias simulator”**. Participants train a mini-AI for sorting applications by choosing the training data (balanced or unbalanced) and observe live how the predictions become unfair. **Animation script:** 1. First pass: balanced data → the AI selects according to skills. Everything is fine. 2. Second pass: unbalanced data (history 90% stars) → the AI ​​begins to favor even less competent stars. 3. Question to the group: *“Has AI changed?” No. What has changed? The data. »* **Follow-up questions:** - “Who chooses the training data?” » (humans — therefore human choices) - “How to correct? » (more diverse data, regular tests, varied teams, audits) #### Important nuance to give Bias can be **corrected** — not perfectly, but a lot: diversify the data, test AI on all groups, have diverse design teams, have systems audited. It is a profession that is developing (see work sequence). --- ### 3.3 Privacy and consent (15 min) #### Where does the data come from? Often… from us! Question to the group: *“In your opinion, what was used to train the AI ​​​​that recognizes faces, writes texts, recommends videos? »* Make the list together: - 📸 **Our photos** published on social networks - ✍️ **Our texts**: posts, comments, opinions, blogs - 🔍 **Our research** on the internet - 👆 **Our clicks and viewing time** (what we watch, for how long, what we don't know) - 📍 **Our location** (phone GPS) - 🎤 Sometimes **our voice** (voice assistants) > 💡 Shock formula: *“If it’s free, your data is often the payment. »* (To be qualified: this is not always true, but it is a good questioning reflex.) #### Consent Key question: **“Have you been asked for permission?” »** - In theory yes: the famous “general conditions of use” which we accept… without reading them. Fun fact: reading them completely for all of our services would take **weeks** per year. - In practice: consent is often vague, hidden, or “all or nothing” (accept or not use the service). #### GDPR — our European shield **GDPR = General Data Protection Regulation** (in force in Europe since 2018). To be presented as a list of **concrete rights**:| Law | In plain language | |---|---| | Right of access | “Show me the data you have on me” | | Right of rectification | “Correct what is wrong” | | Right to erasure | “Delete my data” (the “right to be forgotten”) | | Explicit consent | I must be asked clearly, not tricked | | Consent of minors | Before 15 years (in France), parental consent required for many services | Companies that fail to comply with GDPR risk huge fines (up to 4% of their global turnover). In France, the data policeman is called the **CNIL** (Commission Nationale de l’Informatique et des Libertés). #### Demo: the privacy checker (web page, tab 2) Open the **“Privacy Checker”** tab. Each participant checks what corresponds to their own phone (apps installed, permissions granted). The page calculates an “exposure profile” and gives concrete advice. **No guilt**: the objective is awareness, not fear. This is also the basis of exercise 2 (home audit). --- ### 3.4 Transparency: the “black box” problem (10 min) #### The image to use > *“Imagine a weird vending machine: you insert your file, a light flashes, and it takes out a “REFUSED” paper. You ask why. The machine does not respond. No one — not even its manufacturers — can explain precisely why. That's what a black box is. »* Why this is the case technically (simple version): large modern AIs make their calculations through millions (or billions) of “internal settings” automatically adjusted during learning. No human has written a rule like “if X then refuse”. The result often works very well... but the precise explanation of *each* decision is very difficult to give. #### Why it’s serious in certain areas Ask the question: *“In what cases do we REALLY need to know why?” »* - 🏥 **Health**: “The AI says you do not need this treatment. " For what ? A doctor should be able to check. - ⚖️ **Justice**: in certain countries, AI helps assess the “risk of recidivism”. Deciding on someone's freedom without explanation? Unacceptable. - 🏦 **Bank**: “Credit refused. » Without reason, it is impossible to challenge or correct your situation. **The principle that results from this** (and that we find in the laws): the more impact a decision has on a human life, the more explanation and human control we must demand. The field of research that tries to open the black box is called “Explainable AI”. --- ### 3.5 Responsibility: who is guilty? (10 mins) #### The mini-blitz debate on the autonomous car Scenario to present: *“An autonomous car, in automatic driving mode, does not detect a pedestrian and causes an accident. Who is responsible? »* Have a show of hands vote between: 1. 🧑 **The “driver”** (he was behind the wheel, should he be watching?) 2. 🏭 **The manufacturer** (it was his software that failed) 3. 💻 **The developers** of the AI (they wrote the code) 4. 🤖 **The AI itself** (trap: an AI is not a legal person — it cannot be judged or punished) 5. 🏛️ **The State** (it allowed these cars on the road) Points to highlight: - **AI cannot be responsible**: no conscience, no assets, no legal personality.Responsibility is always *human or corporate*. - The answer depends on the **level of autonomy** promised: if the manufacturer says “keep your hands on the wheel”, the driver retains some responsibility; if he sells total autonomy, the responsibility shifts to him. - This is a legal project **in progress** throughout the world ⚠ — the laws are being written as we speak. Your participants will one day vote on these topics! > 💡 Simple parallel: if a dog bites someone, we don't judge the dog — we turn to the owner. For AI: we turn to those who design, sell and use it. --- ### 3.6 Environmental impact (5–7 min) ⚠ Order of magnitude figures (to be checked before the session, they change quickly ⚠): - **Training** a very large AI model can consume as much electricity as **hundreds of homes for a year**, with CO2 emissions comparable to **several transatlantic flights** (some estimates speak of tens to hundreds of tons of CO2 for a single training). ⚠ - **Using** AI also counts: a query to a large generative model consumes significantly more energy than a traditional web search. Multiplied by billions of requests per day… ⚠ - **Water**: data centers must be cooled, often with water. ⚠ To be nuanced (always both sides): - Actors are working on more **sober** models and data centers powered by renewable energies. ⚠ - AI can also **help** the environment: optimize electricity networks, predict the weather, monitor deforestation, improve agriculture. **Discussion question:** “Is it worth it to use big AI to generate a funny meme?” And to help discover a medicine? Where do we put the cursor? » (No correct answer — it's a trade-off.) --- ### 3.7 Work and automation (5–7 min) Structure in three columns on the board: | 🔄 Transformed professions | ⚠️ Endangered Tasks | ✨ Professions that appear | |---|---|---| | Doctor (helped with diagnosis) | Repetitive data entry | Specialist in AI instructions (“prompt”) ⚠ | | Graphic designer (generative tools) | basic translation | Algorithm auditor / bias tester | | Prof (custom supports) | Standard document sorting | Data Protection Officer (job created by the GDPR!) | | Developer (code wizards) | Very simple customer responses | AI Ethicist | Key messages: 1. History repeats itself: the computer has made professions disappear (typist) and created many more (all digital). AI will likely follow a comparable path — but **faster**, making continued education crucial. 2. The right formula: *“AI is probably not going to take your job. But someone who knows how to use AI could transform it. »* Hence the interest… of this course 😉. 3. **AI as a tool, not as a replacement**: in most professions, AI removes repetitive tasks and leaves judgment, relationships, creativity and responsibility to humans. --- ### 3.8 Rules and laws: the AI Act (10 min) ⚠ In **2024**, Europe adopted the first major law in the world specifically on AI: the **AI Act** (“AI law”). Its application is gradually deployed over several years ⚠. Brilliant idea to remember: classify uses of AI by **risk level** — like a pyramid.#### The risk pyramid (to draw on the board) | Level | Examples | Rule | |---|---|---| | 🔴 **Unacceptable** | Widespread social scoring of citizens, manipulation of vulnerable people, some mass biometric surveillance | **INTERDIT** | | 🟠 **High risk** | AI for recruitment, bank credit, justice, exams, medical devices | **Authorized under strict conditions**: tests, transparency, human control, documentation | | 🟡 **Limited risk** | Chatbots, generated images, deepfakes | **Obligation of transparency**: clearly say that it is AI / generated content | | 🟢 **Minimal risk** | Spam filter, video game AI, basic recommendations | **Free** (good practices encouraged) | Memory tip: **the more AI touches people's lives (freedom, money, health, education), the higher it rises in the pyramid.** Guess: Give examples and ask the group to classify them. (“An AI that grades your exam papers?” → high risk. “An AI that offers you a playlist?” → minimal. “A government that gives good citizen points?” → unacceptable.) Reminder of the European regulatory duo: **GDPR** = protects *data*; **AI Act** = regulates the *uses of AI*. They complement each other. --- ### 3.9 Key activity: the AI Tribunal (20 min) **Objective:** to experience a real structured ethical debate, to argue an imposed position (not necessarily one's own — it's the best exercise in critical thinking that exists). **The case to be judged (recommended):** *“The Jules-Verne college wants to use an AI to grade papers and predict student results. Should it be authorized? »* (Alternative cases in the **“Dilemmas”** tab of the web page: facial recognition cameras in college, medical AI without a doctor, virtual friend chatbot, etc.) #### Organization (court structure) 1. **Distribute roles** (2 min): - ⚖️ One or three **judges** (including the teacher if young group) - 🟢 Team **defense** (for scoring AI): 3–5 people - 🔴 Team **accusation** (against): 3–5 people - 👥 The rest: **jury** (vote at the end) 2. **Preparation** (5 min): each team prepares 3 arguments. Distribute the help sheets (exercise 3). 3. **Pleadings** (2 min per team, timed): defense, then accusation. 4. **Right of response** (1 min each). 5. **Jury questions** (3 min). 6. **Jury vote** + **reasoned verdict** of the judges (3 min). Use the “Dilemmas” tab on the web page to vote and view the results. 7. **Debrief** (3 min) — the most important: *“What was difficult? Have you changed your mind? What guarantees would be required for this to be acceptable? »* #### Arguments that teams can come up with (to help you restart) **Defense (for)**: faster grading, same criteria for everyone (no tired teachers or customer heads), early detection of students in difficulty, teachers released to support. **Accusation (against)**: possible bias (the AI ​​learns old grades… already biased?), black box (how to challenge your grade?), sensitive children's data (GDPR!), prediction = label that confines (“the AI ​​said you will fail”), the AI ​​Act classifies education as **high risk**. **The final twist to be revealed in the debrief:** this case is not science fiction. Automatic scoring systems have already been used — and contested.In the United Kingdom in 2020, an algorithm awarded grades for canceled final exams (Covid): it disadvantaged students from disadvantaged high schools, provoked demonstrations (“F*** the algorithm!”) and ended up abandoned. Real life has already judged this trial. --- ### 3.10 Summary and quiz (5 min) The 5 ideas to be reformulated by the participants: 1. An AI is never more accurate than its data (**bias**). 2. Our data is worth gold; the **GDPR** gives us rights over it. 3. The more an AI decision impacts a life, the more we must demand **transparency** and human control. 4. Responsibility is always **human** — never that of AI. 5. The **AI Act** classifies AI by risk level: from prohibited 🔴 to free 🟢. Quiz (10 questions) in class or at home. Fixed commented in `quiz/quiz.md`. --- ## 4. Materials and preparation - [ ] Video projector + `slides/slides.md` - [ ] `webpage/index.html` webpage tested (works **offline**, no account required) - [ ] Court files printed (exercise 3) or posted - [ ] ⚠ 15 min watch: status of the AI Act, recent energy figures, fresh news example (there's always one!) - [ ] A local/recent news example for the teaser if possible ## 5. Adaptations - **Public 12–14 years old:** favor the school/telephone/social media examples; simplify the AI ​​Act to “prohibited/highly monitored/must prevent/free”; Court in short version (1 min pleadings). - **Adult public:** dig into GDPR (concrete rights, CNIL, how to exercise a right of access), AI at work (employee rights, AI and HR = high risk), and regulatory news ⚠. - **1:30 a.m.:** cut the autonomous car mini-debate (mention it in 2 mins) and reduce the Tribunal to 15 mins. - **2:30 a.m.:** do two trials (two different cases) by reversing the roles — spectacular to show that we can argue both sides. ## 6. Pitfalls and difficult questions | Trick question | Suggested answer | |---|---| | “So AI is racist/sexist? » | AI has no opinions or intentions. It *reflects* the biases present in its data and among its designers. The right word is “bias”, and that can be measured and corrected — it’s our responsibility. | | “What good is GDPR if the giants do what they want? » | They have already paid fines of several hundred million, even billions of euros ⚠, and changed their practices in Europe. Imperfect, but far from useless — Europe has inspired similar laws elsewhere. | | “AI is going to take all the jobs, right? » | Nobody knows for sure. The history of technology suggests transformation rather than disappearance — but faster this time. The best protection: understanding and knowing how to use these tools. | | “Why not ban everything, that would be simpler? » | Banning everything also means giving up the benefits (health, accessibility, science) and letting other countries decide the rules for us. The whole issue is sorting: this is the logic of the AI ​​Act pyramid. | | “Can an AI go to prison? » | No. No legal personality, no conscience, no body. We judge the humans and companies behind it. | ## 7. To go further (teacher)- Documentary *Coded Bias* (on the work of Joy Buolamwini) — FR subtitles available. - **CNIL** website: free educational resources on personal data, very suitable for young people. - Text of the AI ​​Act popularized (official summaries of the European Commission). ⚠ - The grades affair in the United Kingdom (2020): search for “Ofqual algorithm 2020” — perfect for preparing the Tribunal debrief.