Today, we discover

1. The test: true or generated?
2. Why does the AI “hallucinate” (and seem so confident)
3. The false card: error, satire, disinformation, deepfake
4. The S.T.O.P. method — check in 4 reflexes
5. Zoom deepfakes: clues… and their limits
6. Activity: you are the editorial staff!
7. The charter of responsible sharing

The test: true or generated?

I will show you 6 images.

For each, vote by show of hands:

• REAL photo
• AI-generated image

No cheating, no zooming on the neighbor's screen.
We count the points at the end!

Test verdict

Almost everyone has made a mistake at least once.
This is normal — and this is the starting point of the session.

Three years ago, we spotted an AI image in one second.
Today, even the experts are hesitant.

The right question is no longer:
“Can I SEE the difference?”

But: “How do I CHECK?”

Remember Session 3…

A language AI does ONE thing:

Predict the most likely outcome of a text.

It does not consult a “base of truths”.
She doesn't know what's true — she knows what's plausible.

Plausible ≠ True ← the most important sentence of the day

The reader parrot

Imagine a parrot that has read the entire library.

• He speaks beautifully
• He knows millions of sentences
• But ask him for a book that doesn't exist...

He will invent one for you — credible title, credible author,
credible publisher. His job: to produce credible.
Not verifying the real.

This is called a hallucination.

Live demo

Let's ask a chatbot:

1. The biography of a invented person
2. “Cite me 3 scientific articles on [special subject]”

Then… let’s check his answers.

Spoiler: at least one risk reference
not to exist. With a perfectly credible title.

3 reasons to stay vigilant

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 answer, rarely to say “I don’t know”

AI does not “lie”: lying requires intention.
She is wrong with confidence — important nuance!

False precision

What makes a hallucination so convincing?

• A specific date: “in 1987”
• A proper name: “Marcel Dupreuil”
• A concrete detail: “at Éditions du Vieux Clocher”

False precision is the favorite costume of the false.

The serious form of an answer proves nothing.

Falsehood has many faces

Question: which one is the most dangerous?

Nuanced response

• Wilful disinformation is the most toxic…
• …but an error shared a million times does just as much damage.

Regardless of the origin:
it is SHARING that gives its power to falsehood.

And satire? No problem… as long as it is identifiable.
The problem: the screenshot circulating without the name of the site.

S.T.O.P. — your 4 reflexes

S — Source

Who publishes? Is the site/account identifiable, old, reliable?

T — Traces

Are other serious media talking about it?

O — Origin

Where do the image, the quote, the number come from?

P — Break

Strong emotion? This is the signal: wait before sharing.

S for Source

Questions to ask yourself:

• Does this media/account really exist?
• Is there an “About” page, identified authors?
• Was the account created 3 days ago?
• Is the site address weird? (sante-choc-info.biz )

Information without a source is a rumor with a nice costume.

T for Traces

Reverification — the number 1 weapon of journalists:

A HUGE scoop picked up by no known media?
Maximum alert.

A major event is always covered by
several independent media.

“No one talks about it except this account” =
red flag, not proof of conspiracy.

O for Origin

For a suspicious image:

• Reverse image search (Google Images, TinEye, Lens)
• We often discover… an old recycled photo
  in a false context!

For a quote:

• Look for it in quotes in a search engine
• Who said it? When ? Or ? In what context?

P for Pause (the most important!)

Disinformation is designed to trigger:

• anger
• fear
• indignation
• enthusiasm (“too good to be true”)

Emotion short-circuits reflection.
Feeling a strong emotion = THE signal to take a break.

Your superpower: 30 seconds wait.

Let's train together

“URGENT College X replaces all its teachers
by AI from September! Share quickly!!!”

• S: unknown site, no “About” page
• T: no media is talking about it — for such a scoop?!
• O: photo from an image bank
• P: fear + urgency + “share quickly”

Verdict: intox. Four red flags out of four.

Classic visual cues

To observe in a suspicious image or video:

• Strange hands and fingers, asymmetrical jewelry
• Illegible texts in the background (signs, posters)
• Inconsistent reflections and shadows
• “Fade” backgrounds, repeated details
• Skin too smooth, teeth too regular

Audio and video clues

• Imperfect lips/sound synchronization
• Absent breathing, flat intonation (cloned voices)
• Rare or mechanical blinks

A few seconds of your voice are enough
today to clone it.

The warning that counts

These clues age FAST.

Recent models correct hands, reflections,
the blinks...

The lasting lesson is not the list of clues.
It's the reflex: spectacular video
→ check the source BEFORE believing it.

Your eyes will betray you one day. Your method, no.

Protect yourself concretely

The cloned voice scam: a “relative” calls,
panicked voice, urgent request for money…

The family code word:
agree on a secret note with your loved ones
to check a suspicious emergency call.

Simple, free, effective.
Talk about it at home this evening!

What about AI detectors?

“There are tools that detect AI content, right?”

• Yes… but they are often wrong, in both directions
• A real text can be accused of being “AI”
• A neat deepfake can get through

No tool replaces source verification.
S.T.O.P. > any detector.

The game of writing

You are the editorial staff of the newspaper “Le Véracteur”.

In teams of 3-4: three dispatches arrive.
For each, decide:

• PUBLISH
• CHECK AGAIN
• REJECT

Justify each verdict with S.T.O.P. — 4 min per dispatch!

The editor-in-chief’s scale

• Publish a mistake: -10 points
  (you are misleading your readers)

• Rejecting real information: -5 points
  (you deprive your readers of the truth)

• Check again: 0 points
  (honest caution… but a newspaper must decide!)

Rejecting the real thing also costs points.
Critical thinking ≠ reject everything!

Debrief: the octopus trap

Was the “unbelievable but true” report rejected?

This is the liar's dividend:

If the public believes that “everything can be faked”,
then liars can deny... even the real evidence.

Critical thinking does not mean rejecting everything.
It’s methodically checked.

The responsible sharing charter

Before sharing, I agree to:

1. Do S.T.O.P. if the information makes me react strongly
2. Check the source before believing a shocking video
3. Do not share when in doubt — doing nothing is already taking action
4. Correct a loved one with kindness (privately)
5. Report misleading content on platforms

Choose ONE commitment and write it down.

The 5 ideas to remember

1. Generative AI produces plausible, not real
2. Falsehood has many faces: error, satire, disinformation, deepfake
3. S.T.O.P.: Source · Traces · Origin · Pause
4. Deepfake indices are getting old; the reflex to check, no
   5.Checking before sharing is protecting others