#Exercises — Session 5 🟢 # “AI that sees, hears and creates” **Program:** Applied AI — Beginner Level · Instructor: Yann Isola **Material:** paper + pencil. Access to a generative AI tool for exercise 3 (“screenless” version provided). --- ## Exercise 1 — Describe the image for a blind AI 👁 **Duration: 15 min · In pairs** ### The idea An image generation AI does not “see” what you have in mind: it only knows the **words** of your description (the *prompt*). This exercise puts you in his shoes…and teaches you how to write good descriptions. ### Unrolled **Step 1 (5 min):** Player A secretly chooses a simple image (provided by the instructor, or an image of his choice: a landscape, a scene, an animal in a setting). Player B does not see it. **Step 2 (5 min):** Player A describes the image **only with words**, as if writing a prompt for an AI. Player B draws what he understands (a quick, ugly drawing is perfect!). Forbidden to look at B's drawing along the way — an AI doesn't get a second chance either. **Step 3 (5 min):** Compare the drawing and the original image. Discuss: - What was missing from the description? (colors? positions? sizes? atmosphere?) - Rewrite together a “version 2” description that would have given a better result. ### What to remember A good description for a precise AI: **the subject** (what?), **the action** (what is he doing?), **the setting** (where?), **the style and atmosphere** (realistic? cartoon? day? night?). The more precise it is, the more the result corresponds to what you imagine. This is exactly the job of “prompting well”. ### Prompt self-assessment grid (to be checked) - [ ] I named the main subject - [ ] I described what it does - [ ] I described the setting/background - [ ] I gave the colors or the atmosphere - [ ] I specified the style (photo, drawing, painting, etc.) --- ## Exercise 2 — The deepfake detective challenge 🕵 **Duration: 15 min · In small groups (3-4)** ### The idea Learn to spot content created by AI — and above all, adopt the right reflexes **even when the image is perfect**. ### Part A — The hunt for clues (5 min) The instructor projects a series of images (or use the **detective game** from the session webpage). For each image, your group votes: **REAL** or **AI**? Note for each image the clue that made your decision: | Picture | Voting (Real / AI) | Decisive clue | |---|---|---| | 1 | | | | 2 | | | | 3 | | | | 4 | | | | 5 | | | | 6 | | | ### Part B — The detective’s checklist (5 min) As a group, classify these clues into two families: 👁 **visual clues** (in the image) and 🧠 **context clues** (around the image): weird hands · anonymous account posting · illegible text in the background · no other source talks about it · inconsistent reflections in the glasses · the image provokes a very strong emotion · strange teeth or ears · publication that is “too good” (election, scandal, etc.) · blurred and distorted background · known and reliable media **Trick question:** In a year, will visual cues still work as well? Which ones will remain valid forever? ⚠ ### Part C — The scenario (5 min) Your group receives this scenario: > *You receive a video on your phone where a celebrity announces something incredible. A friend urges you: “Share, quickly!! »*Write in 3 lines what you do, in order, **before** sharing (or not). ### What to remember The visual defects of deepfakes disappear as AI advances ⚠. The reflexes that remain valid forever: **check the source, cross-check information, be wary of strong emotions**. A modern detective checks *around* the picture as much as *in* the picture. --- ## Exercise 3 — My mini AI studio 🎨🎵 **Duration: 25-30 min · Alone or in pairs · Creative project** ### The idea You become a mini-creation studio: you will produce a small work by combining **at least two contents** created (or designed) with AI, then present it. ### Mission Choose a theme from (or invent your own, validated by the instructor): - 🚀 “A postcard from the year 3000” - 🐉 “The fantastic animal of our class” - 🌍 “A poster to protect the planet” - 🎬 “The trailer for a film that doesn’t exist” ### Version A — With access to AI tools ⚠ 1. **Image:** write a neat prompt (use the grid from exercise 1) and generate an image with the tool indicated by the instructor. You are allowed **3 attempts maximum** — improve your prompt between each attempt and keep track of your 3 versions. 2. **Text:** Ask a text AI for a slogan, mini-poem or 3-sentence synopsis to accompany your image. 3. **Sound bonus:** if a music tool is available ⚠, describe the background music of your work (style, instruments, tempo) and generate 30 seconds. 4. **Assemble** everything (one page, one slide or simply side by side). ### Version B — Without screen (just as valid!) 1. **Image:** write the perfect prompt for your image (5 lines max) and **draw** quickly what you think the AI would produce — including its likely errors (hands? text? details?). 2. **Text:** write the slogan yourself, then write the prompt you would have given the AI ​​to get it. 3. **Sound:** describe in 3 lines the music you would ask of a musical AI (style, instruments, atmosphere, tempo). ### Presentation (1 minute per studio) Imposed structure: 1. **What I asked** (my prompt) 2. **What I got** (show) 3. **What I had to correct** (the gap between the idea and the result) ### Golden rules of the studio 📜 - ❌ Never the face, name or voice of a **real person** without their consent (classmates, teachers, celebrities: prohibited). - ✅ We always say it's made with AI — an honest creator doesn't pass AI off as his own hand. - ✅ The AI ​​suggests, **you** decide: the choice, sorting and corrections are your job as a creator. ### What to remember Creating with AI is not about pressing a button: it is **describing precisely, evaluating the result, correcting, starting again**. The idea, the taste and the responsibility remain human. --- ## To go further (at home, optional) - Activate **automatic subtitles** on a YouTube video and count the errors over 2 minutes: voice recognition is good, but not perfect. What words is she wrong? (proper names, technical words, when speaking quickly, etc.) - Ask a loved one if they know about **cloned voice** scams — explain to them how to protect themselves from them (call the person back on their real number, agree on a code word as a family).