Français
Applied AI · Beginner 🟢 · Session 5
📝 Teacher's Guide
← Return to program 📄 Source .md

Teacher's Guide — Session 5 🟢

“AI that sees, hears and creates”

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 4 (concepts: AI, machine learning, data, text generative AI)

1. Educational objectives

At the end of the session, each participant should be able to:

  1. Explain what is an AI multimodal with your own words (an AI that processes several types of content: text, image, sound, video).
  2. Describe in 3 simple steps how an AI “sees” an image (pixels → visual cues → recognition).
  3. Quote at least 3 concrete applications of computer vision and 2 of voice recognition.
  4. Appoint the major generation tools (images, music, video) and understand that they create from a text description. ⚠
  5. Define what a deepfake is, explain why it’s a problem, and list 4 clues to spot one.
  6. Identify positive applications: accessibility, medicine, artistic creation.
  7. Create a mini-project combining several AI-generated contents (practical activity).

⚠ Volatile content : the names of tools, their capabilities and their prices evolve very quickly. Check the state of the art before each session (15 minutes of standby time is enough). The sections marked ⚠ in this guide are the ones to double-check.

2. Overview and common thread

Common theme of the session: “So far, we have spoken to the AI ​​with text. Today, we discover that she also has eyes, ears, a voice... and a paintbrush. »

The session alternates short presentations, demos, discussions and a large final creative activity. The highlight is the deepfake detective game (interactive web page): it creates emotion and anchors the message of caution.

Timed course (based on 2 hours)

Time Sequence Format
0:00 – 0:10 Tagline: “Does AI have senses? » Discussion
0:10 – 0:25 Multimodal AI: definition + examples Presentation + demo
0:25 – 0:45 How AI “sees”: pixels → features → recognition Lecture + pixel explorer (web page)
0:45 – 1:00 Image generation ⚠ + voice recognition and synthesis Presentation + demos
1:00 – 1:10 Break —
1:10 – 1:25 Music, video, and deepfakes Presentation + detective game (web page)
1:25 – 1:35 Positive applications: accessibility, medicine, art Presentation + discussion
1:35 – 1:55 Activity: Create with AI (exercise 3) Workshop
1:55 – 2:00 Summary + quiz (or quiz at home) Quiz

3. Detailed content, sequence by sequence

3.1 Hook (10 min) — “Does AI have senses? »

Ask the group: “Have you ever used an AI that doesn’t read text, but sees or listens? »

Expected answers (bring them out, don’t give them):

Key message: Multimodal AI is not science fiction, everyone in the room is already using it, often without knowing it.

3.2 Multimodal AI (15 min)

Simple definition: “multi” = several, “modal” = content types (modes). A multimodal AI can work with text, images, sound and video — as input (it understands them) and/or output (it creates them).

Effective analogy: “Early text AIs were like texting someone in the dark. Multimodal AI is the same person but who can now look at your photos, listen to your voice and respond to you by drawing. »

Table to build on the board with the group:

Human sense AI equivalent Everyday example
The view 👁 Computer vision Face unlock
Hearing 👂 Voice recognition “Hey Siri…”
The word 🗣 Speech synthesis Talking GPS
The drawing ✍ Image generation DALL-E, Midjourney ⚠

Important nuance to give: AI doesn’t “see” or “hear” like we do. It transforms everything into Numbers and looks for patterns in these numbers. No consciousness, no feeling — calculation. This is the bridge with the following sequence.

3.3 How AI “sees” (20 min) — technical heart of the session

This is the most “technical” passage: stay very visual. Use pixel explorer of the web page.

Step 1 — Pixels. A digital image is a grid of small colored squares: the pixels (from English picture element , “image element”). Each pixel is described by 3 numbers: the amount of red, green and blue (the code RGB , in English RGB — Red, Green, Blue ), each from 0 to 255. For the computer, a photo of a cat = a huge table of numbers. Nothing else.

Demonstration with Pixel Explorer: Gradually zoom in on the image until you see the individual squares and their RGB values.

Step 2 — Visual cues (features). The AI ​​learns to spot patterns in these numbers: first very simple things (outlines, lines, spots of color), then by combining: shapes (an eye, a pointy ear, whiskers), then whole objects (a cat!). In English we call these indices features (" features "). It’s a stack: simple → medium → complex.

Analogy: “It’s like learning to read: first the letters, then the syllables, then the words, then the sentences. The AI ​​does the same with images: lines → shapes → objects. »

Step 3 — Recognition. After seeing millions of labeled images (“this is a cat”, “this is a dog”), the AI ​​knows how to associate the detected patterns with a response, with a confidence score : “97% cat”. Reminder from session 2: this is learning from examples, not magic.

Concrete applications to be detailed:

Limit to mention: the vision AI can make surprising mistakes (a sticker on a STOP sign can confuse it). She recognizes patterns, she doesn’t “understand” the scene like we do.

3.4 Image generation (8 min) ⚠

Principle: you write a description (a prompt ), the AI ​​draws the image. She learned by analyzing millions of image + description pairs, and she learned to go the opposite way: from text to image.

Popularization of operation (optional, for the curious): many of these AIs work by broadcast : they start from a “fog” of random pixels and “denoise” it little by little, guided by your description, until the image appears. Like a Polaroid that reveals itself, but driven by the text.

Tools to cite ⚠ (check before the session):

Points of vigilance to convey:

3.5 Speech recognition and speech synthesis (7 min)

Voice recognition (the machine listens): your voice = a sound wave = more numbers. The AI ​​has learned, over thousands of hours of speaking, to associate these waves with words. Simple chain to draw on the board: sound → numbers → text .

Applications: voice assistants (Siri, Alexa, Google Assistant), voice dictation, and above all automatic subtitles — a small revolution: YouTube, videoconferencing, live TV... access to content for deaf or hard of hearing people, and for all those who watch without sound.

Speech synthesis (the machine speaks): the reverse path, text → sound . In English : text-to-speech (TTS, “text to speech”). ⚠ Current voices have become very natural: intonations, pauses, emotions. Some tools can clone a voice from a few seconds of recording – which makes the direct link with deepfakes (sequence 3.7).

Applications: GPS, reading articles and audio books, voice for people who cannot speak (the historical example of Stephen Hawking speaks to all ages).

3.6 Music and video: current frontiers (7 min) ⚠

Section very volatile — to re-check systematically.

📌 State-of-the-art benchmark (July 2026, to be re-checked ⚠): the large general models are now divided into families. On the OpenAI side, GPT-5.6 (July 9, 2026) declines Sol/Terra/Luna And Sol Ultra : reasoning effort until max , and an “ultra mode” which mobilizes sub-agents; Earth roughly aims for the GPT-5.5 level at half price. xAI side, Grok 4.5 (July 8, 2026) is positioned “Opus-class”, very strong in code and agentic, with API at approximately $2/M entry tokens And $6/M exit tokens , and an effort of reasoning low / medium / high ; xAI also acquired Cursor. Meta side, Muse Spark 1.1 (Meta Superintelligence Labs, July 9, 2026) becomes the multimodal and agentic-coding successor to Llama; Spark 1.0 dated April 2026, Muse Image of July 7, and the Meta Model API is in preview. Don't memorize version numbers — they change every month. Remember the logic: each publisher offers a large model powerful and expensive, and lighter variants cheaper for simple tasks. It’s this “right model for the right task” reasoning that matters, not the name of the day.

Key message: what is “impossible” today may be commonplace in a year. Hence the importance of learning to check what we see — perfect transition to deepfakes.

3.6-bis World models — the next frontier (7 min) ⚠

🎯 Why this sequence? Participants have just seen an AI write, draw, speak, film. Natural question: “does she understand really the world? » This is the subject of models of the world (in English world models ), one of the major research projects in 2026. Volatile section but the basic idea is sustainable.

The starting point to ask in the room: a text model, however impressive it may be, lives in a universe of words . He has read billions of sentences, but he has never reside a world. Little riddle to read aloud to show:

“I'm going to wash my car; the car wash is 1 km away. How many kilometers do I travel on foot ? »

Many AIs (and distracted humans!) respond “1 km” by associating words. The correct answer is zero : to wash your car, we go there by car . To answer correctly, it is not enough to manipulate words — it is necessary reconstruct the scene : what a car is, a station, the act of driving, the purpose of the action. This is exactly what a pure language model lacks.

The simple definition (to be noted): A model of the world , it is a internal representation of how the world works , which allows anticipate the consequences of an action . If I push this glass, will it slide, roll or fall? If I turn the steering wheel, where does the car go?

The analogy that works for all ages — the baby and the cat:

→ Have participants say: our understanding of the world is older than our words. When you go down a flight of stairs, you don't calculate gravity: your body simulated balance and movement.

Why it matters (the central message): as long as we stay in the text , the error is tolerable — a vague sentence can be corrected, an invented source can be verified. But as soon as we move on to physical world , tolerance for error collapses. Striking example to tell: a robot who has to catch a glass cannot produce a “plausible” answer — he must know the exact position of the glass, the force to be applied, the fragility of the object. If he is wrong, he is not making a grammatical error: he breaks the glass .

The image → video → robotics link (to be drawn like a staircase):

  1. A picture already gives a structure of space: objects, positions, depths.
  2. A video add the time : things move, fall, bounce, disappear then reappear.
  3. There robotics add the action : the model is no longer content to observe, it must try, make a mistake, correct and live with the consequences.

It is in this meeting between language + vision + video + action that many researchers situate the next generation of AI: no longer just “what word comes next?” ", but " If I do this, what will happen? » We also talk about spatial intelligence .

Final message of the sequence (to hammer home, without jargon): today’s AI knows how to describe the world; the models of the world seek to do so anticipate . It is perhaps the “missing piece” between a brilliant chatbot and an AI that acts in real life (cars, domestic robots, physical assistants).

⚠ Safeguards for the trainer: remain mainstream. No formalism. If a participant asks “does this already exist?” »: yes, in the form of research prototypes and video generation models which begin to respect (imperfectly) physics – hence the “objects which appear/disappear” seen in sequence 3.6. The subject is young : Present it as direction, not a finished product.

3.7 Deepfakes (8 min) — critical moment of the session

Definition : A deepfake (of deep learning , “deep learning”, + fake , “false”) is content – ​​image, voice or video – created or modified by AI to make people believe in something false: making a person say words that they have never said, placing them in a scene that never existed.

Why this is a problem:

How to spot them – the detective’s 4 reflexes:

  1. Visual details: hands (too many or fewer fingers), teeth, ears, asymmetrical jewelry, illegible text in the background, inconsistent shadows and reflections. ⚠ These defects diminish over time — emphasize reflexes 3 and 4 which remain valid.
  2. The movement (video): strange blinking, lips poorly synchronized with sound, facial contours that “float”.
  3. The source : who published? A reliable media or an anonymous account created yesterday? Does the information exist elsewhere?
  4. Common sense: is this likely? Shocking content that is “too good” should arouse doubt. A strong emotion (anger, fear, indignation) is the signal that must be checked before sharing.

Integrated activity: the deepfake detective game on the web page (real or AI?). Allow 5 minutes, play in teams, debrief the clues.

Final message to hammer home: the goal is not to become paranoid, it is to become auditor . To doubt ≠ to no longer believe anything. Doubt = check before believing and especially before sharing.

3.8 Positive applications (10 min)

It’s essential to end on the positive after deepfakes — multimodal AI is changing lives:

3.9 Final activity: create with AI (20 min)

See exercices.md, exercise 3 (creative mini-project). In summary: each participant or pair creates a mini-work combining at least two AI-generated contents (e.g.: an image + a text, or an image + a musical idea), then presents in 1 minute: what I asked → what I got → what I had to correct .

If no access to AI tools in the room: use the “paper” version described in the exercise (write the prompts and predict the results, then compare with examples prepared by the instructor), or the creative gallery on the web page.

3.10 Summary and quiz (5 min)

The 6 ideas to remember (have them reformulated by the participants):

  1. Multimodal AI sees, listens, speaks and creates — not just text.
  2. For AI, everything is numbers: pixels, sound waves. It detects patterns, it does not “feel” anything.
  3. Computer vision: pixels → visual cues → recognition.
  4. Image/music/video generation: a text description is enough — the tools are evolving at full speed. ⚠
  5. Deepfakes: check the source, keep a critical mind, do not share out of emotion.
  6. Models of the world: describe the world (text) ≠ anticipate it (act). The next frontier connects image, video and action — like a baby who understands space before speaking. ⚠

Quiz: 12 multiple choice questions (see quiz.md) — in class or at home.

4. Materials and preparation

5. Teaching pitfalls and difficult questions

Situation Recommended response
“So the AI ​​really sees? » No — it calculates on numbers and detects patterns. No consciousness, no visual experience. The word “see” is a convenient image.
“Are deepfakes illegal? » Creating a deepfake is not always illegal (assumed parody, cinema); using it to deceive, scam, harass or create intimate content without consent is. Do not play lawyer: give the main principles.
“AI will replace artists? » Open question. History shows that new tools (photo, synthesizer) have transformed art without killing it. Recognize real concerns (illustration professions) without catastrophism or denial.
A participant wants to create a deepfake “for fun” Frame immediately: never with the face or voice of a real person without their consent. This is the absolute rule, including “just for fun”.
“How can you be 100% sure that an image is real? » We can't always. Hence the importance of the source and the cross-checking – it is the skill of the century.
Age gaps in the group 12-15 year olds shine at the detective game; adults hang on medicine/voice scams. Use both angles.

6. Links with other sessions

Have a good session! 🎨🎙👁