Français
Applied AI · Beginner 🟢 · Session 9
📝 Teacher's Guide
← Return to program 📄 Source .md

Teacher's Guide — Session 9 🟢

“Info or fake? Critical thinking in the age of AI »

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 8 (notably S5: multimodal AI/deepfakes discussed, S6: ethics and bias)

1. Educational objectives

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

  1. Explain why generative AI can confidently produce false information (“hallucinations”).
  2. Distinguish the main types of false information: unintentional error, satire, intentional disinformation, deepfake.
  3. Apply a simple verification method in 4 reflexes (method STOP.) when faced with questionable information or image.
  4. Spot the classic clues of a deepfake (image, audio, video) – knowing that these clues are becoming less reliable. ⚠
  5. Adopt responsible sharing reflexes: check before sharing, cite your sources, report.

🧭 Common theme of the session: “AI makes fakes easy to make. Our best defense is not technical: it is a mental reflex. »

What this session IS NOT: a course of fear. We don’t want participants to come away thinking “everything is false, we can’t believe anyone”. The message is: the true exists, we just have to learn to find it. ⚠

2. Overview and common thread

The session alternates demonstrations, detection games and practical methods. It builds on what participants already know:

What’s new in S9: we move from “the AI ​​can make mistakes” to “ how do I check ».

Timed course (based on 2 hours)

Hour Duration Sequence
0:00 10 mins Hook: the “True or generated?” test »
0:10 20 mins Why AI “hallucinates” (and why it seems so confident)
0:30 15 mins The fake map: error, satire, disinformation, deepfake
0:45 10 mins Break ☕
0:55 20 mins The S.T.O.P. method — check in 4 reflexes
1:15 15 mins Zoom deepfakes: visual and audio clues ⚠
1:30 20 mins Key activity: “Writing” (team fact-checking game)
1:50 10 mins Summary, responsible sharing charter, quiz

3. Detailed content, sequence by sequence

3.1 Hook (10 min) — “True or generated? »

Unfolded: project 6 images (prepared in advance, see §4): 3 real photos, 3 images generated by AI. Show of hands for each: “real or AI?” ".

Desired effect: the class makes a mistake at least once — it’s almost guaranteed. This is the trigger for the session.

Key phrase: “Three years ago, we spotted an AI image in one second. Today, even the experts are hesitant. The question is no longer “can I see the difference?” but “how do I check?” »

⚠ Point of attention: do NOT choose shocking or political images. Favor: landscapes, animals, neutral portraits, food. The goal is the mechanism, not the controversy.

Without a screen: print the 6 images in color (A5 minimum) and pass them around.

3.2 Why AI “hallucinates” (20 min)

The heart of the sequence: link this phenomenon to S3. A language AI predicts the most likely outcome , it does not consult a basis of truths.

Analogy of the reading parrot: “Imagine a parrot that has read the entire library. He speaks beautifully. But if you ask him for a book that doesn't exist, he will invent one for you - with a credible title, a credible author... because his job is to produce something credible, not to verify that it is true. »

Live demonstration (if screen available): ask a chatbot:

3 simple reasons to remember (on the board):

  1. She predicts, she doesn't know — plausible ≠ true.
  2. She has a cutoff date — his knowledge stops somewhere.
  3. She wants to please you — trained to give an answer, rarely to say “I don’t know.” ⚠

⚠ Common trap: a participant will say “so the AI ​​is lying”. Correct: lie assumes an intention. AI has no intention; it produces the plausible. The right word is “she is confidently wrong.” This nuance is important so as not to humanize the machine (recall from S1).

Answer key to the flash question (“Why does AI invent sources?”): because an invented source looks like statistically to a real source, and that the model generates what looks like what is expected.

3.3 The forgery card (15 min)

Objective: to show that “fake news” covers very different realities. Draw 4 boxes on the board:

Kind Intention Example
Error None (unintentional) A journalist gets the number wrong
Satire/parody Make people laugh (assumed) Gorafi, April Fools
Disinformation Willfully deceive False quote attributed to a celebrity
Deepfake Deceive with AI Faked video of a personality

Question to the class: “Which is more dangerous?” » Leave the discussion for 3 minutes. Nuanced answer: disinformation voluntary is the most dangerous, but the error shared massively does so much damage. → transition: “It doesn’t matter the origin: it’s sharing that gives the fake its power. »

⚠ Make it clear that satire is NOT a problem as long as it is identifiable. The problem arises when a screenshot of Gorafi circulates without the name of the site.

3.4 Break (10 min) ☕

3.5 The S.T.O.P. method. (20 mins)

The practical deliverable of the session. Write BIG on the board:

Guided micro-exercise: project fake training information (provided in exercises.md, ex. “College X is replacing all its teachers with AIs from September”). Unroll S.T.O.P. together, step by step, writing down the answers.

Answer key to the micro-exercise:

Key phrase: “The P is the most important. Misinformation is designed to short-circuit your brain with emotion. Your superpower is 30 seconds of break. »

3.6 Zoom deepfakes (15 min) ⚠

S5 reminder: we know that AI generates images, voices, videos. Here we learn to observe .

Classic visual cues (to be put into perspective):

Audio/video cues:

⚠ MANDATORY warning to be given: these clues date quickly. Recent models correct hands, reflections, blinks. The lasting lesson is not the list of clues, it’s the reflex: a spectacular video = check the source before believing it. Insist: “Your eyes will betray you one day; your method, no. »

Concrete case to tell: a fake audio recording of a school principal or politician can be fabricated in minutes with just a few seconds of voice. Hence the importance of the “family code word”: some families agree on a secret word to verify a suspicious emergency call (cloned voice scams). It’s concrete, protective, and the participants talk about it at home.

3.7 Key activity: “Writing” (20 min)

Principle: the class becomes the writing of a newspaper. “Dispatches” arrive (prepared maps, see exercises.md, exercise 3). Each team (3-4 people) must decide in 4 minutes per dispatch: PUBLISH / CHECK AGAIN / REJECT , justifying with S.T.O.P.

Unfolded:

  1. Form teams, distribute the S.T.O.P. sheet. (one per team).
  2. Distribute 3 dispatches (mixture: 1 real banal, 1 fake attractive, 1 real but incredible).
  3. Pooling: each team announces its verdicts, the teacher reveals the truth.

Desired effect: the “real but unbelievable” is often wrongly dismissed → excellent discussion: critical thinking does not mean rejecting everything, it means knowing how to verify. ⚠ This is THE subtle point of the session: systematic skepticism is also an error (we call it the “liar’s dividend”: if everything can be false, liars can deny even real evidence).

Detailed corrections to dispatches: see exercises.md, exercise 3 (corrected section).

3.8 Summary and quiz (10 min)

The 5 ideas to remember (to be reformulated by the participants):

  1. Generative AI produces plausible , not of TRUE .
  2. Falsehood has many faces: error, satire, disinformation, deepfake.
  3. STOP. : Source, Traces, Origin, Break.
  4. Deepfake clues are getting old; the verification reflex, no.
  5. Not sharing is already taking action. Checking before sharing helps protect others.

End with the responsible sharing charter (webpage/index.html, final section): everyone chooses ONE personal commitment and writes it down.

Bridge to session 10: “The next session is the last at the beginner level: we bring everything together – final project, assessment, and the rest of your adventure with AI. »

4. Materials and preparation

To prepare before the session:

100% screenless version: the entire session works with the printed images, dispatch cards and board. The hallucination demo is replaced by printed screenshots or by the dramatized reading of a chatbot dialogue (provided in the appendix of exercises.md).

5. Adaptations

6. Pitfalls and difficult questions ⚠

“There are AI detectors, right? » — Yes, but they are often wrong, in both directions. No tool replaces source verification. Never promise that a tool will “detect” for sure.

“Then we can no longer believe anything!” » — Key answer: yes, precisely. Serious media, cross-checking, primary sources still exist. The goal is trust methodical , not general distrust. (See “liar’s dividend”, §3.7.)

“Is it illegal to make a deepfake? » — In France/EU: using someone's image or voice without consent to deceive can be punished (image damage, identity theft); the AI ​​Act requires reporting generated content. Clearly identified satire remains permitted. Stay general, you are not a lawyer. ⚠

“What if a friend shares a poison? » — Advise kindness: private message rather than humiliating public correction. We correct the information, not the person.

Sensitive subject: if a participant brings a divisive political or religious example, refocus on the METHOD (“very good, let’s apply S.T.O.P.”), never on the substance of the debate. ⚠

7. To go further (teacher)

Guide written for the Applied AI program — Yann Isola. Session 9/10 of the Beginner level 🟢.