On the program today

1. Bias: when AI is unjust unintentionally
2. Privacy: where does the data come from? (often from us!)
3. consent and GDPR
4. The black box: why AI cannot be explained
5. Responsibility: who is guilty?
6. The environment and the work
7. The laws: the European AI Act
8. The AI Tribunal — it’s up to you to plead!

A true story to start

A large company is building an AI to sort CVs and identify the best candidates.

A few months later, we discovered that the AI discarded almost all women's CVs.

No one asked him to do this.

So… where does the problem come from?

The answer: data

• The company had recruited mainly men for 10 years
• The AI learned about these past recruitments
• She concluded: “good candidate = male profile”
• She even learned to penalize words like “women’s team…”

The phrase to remember from the whole session:
“An AI is never more accurate than the data that fed it.”

The bias mechanism

The apprentice cook analogy

An apprentice who learns by watching ONE chef…
If the boss salts too much → the apprentice salts too much. Without asking any questions.

Imperfect world → imperfect data → AI that learns these flaws
              → unfair decisions… at LARGE scale and VERY fast

Bias is not a failure: the AI ​​reproduces faithfully what it has been shown. That's the trap.

Very real biases — 3 examples

1. Facial recognition
Less than 1% error on some faces… more than 30% on others (dark-skinned women). Cause: training photos not diverse enough. Consequence: innocent people arrested by mistake.

2. Recruitment
The CV sorter of history — abandoned by the company.

3. Targeted advertising
Well-paid job offers shown more often to men; expensive loans targeting certain neighborhoods.

Demo: the bias simulator

It's up to you! (session web page, tab 1)

1. Train a mini sorting AI with balanced data → everything is fine
2. Train it with unbalanced data → it becomes unfair
3. Has AI changed? No. What has changed? The data.

How to correct bias?
More diverse data · Tests on all groups · Diverse teams · Regular audits

Privacy: where does the data come from?

The AIs learned thanks to… us:

Our published photos	Our research
Our texts (posts, opinions, comments)	Our clicks and viewing times
Our position GPS	Sometimes our voice

“If it's free, your data is often the payment.”
(Not always true — but always worth wondering!)

Consent

Have you been asked for permission?

• In theory yes: the general conditions of use…
• …that no one reads! Reading them all would take weeks per year
• Often vague, hidden, or “take it or leave it” consent

The right question to always ask yourself:

“Does this app really NEED this data to work?”
A maps app that wants your position: logical.
A puzzle game that wants your contacts and your microphone:

GDPR: our European shield

GDPR = General Data Protection Regulation (Europe, 2018)

Fines of up to 4% of global turnover — In France, the police: the CNIL

Demo: the privacy checker

Web page, tab 2 — and exercise 2 at home!

• Check the apps and permissions of your phone
• Discover your “exposure profile”
• Receive concrete advice to regain control

Goal: not fear — lucidity.
We observe → we understand → we adjust.

The black box

Imagine a strange distributor: you insert your file…
beep beep … he takes out a paper: “REFUSED”.
Why? The machine does not respond. No one can explain precisely.

Why? Big AI calculates via millions of internal settings automatically adjusted. No human has written “if X then refuse”.

It often works very well... but explaining each decision is very difficult.

When the black box becomes serious | Domain | The annoying question |

|---|---|
| Health | “No treatment for you. " - For what ? A doctor must be able to check! |
| Justice | Evaluate a “risk of recidivism” without explanation? And the freedom of the person? |
| Bank | “Credit refused.” For no reason → impossible to challenge or improve |

The principle: the more a decision impacts a human life,
the more we must demand explanation + human control.

Responsibility: the crash test

An autonomous car does not detect a pedestrian and causes an accident.
Who is responsible? Vote!

1. The “driver”?
2. The manufacturer?
3. Developers?
4. AI itself?
5. The state that authorized these cars?

The answer (spoiler: not the AI)

• AI cannot be responsible: no conscience, no legal personality. You don't put software in prison.
• Parallel: if a dog bites, we don't judge the dog — we turn to the master.
• The answer depends on the promises of the manufacturer: “keep your hands on the wheel” ≠ “sleep peacefully”.
• Laws are being written right now, all over the world. You will vote on these topics one day!

Responsibility is always human (or corporate). Never that of the machine.

Environmental impact

Training a very large AI means:

• As much electricity as hundreds of homes for a year
• A CO2 footprint comparable to several transatlantic flights
• water to cool data centers

And every use counts: a request to a large generative AI consumes significantly more than a traditional web search… × billions per day

The environment: both sides

The honest debate:

Concerns	Hopes
Exploding consumption	More sober models in development
Water and electricity for data centers	Centers powered by renewable
Futile use = waste?	AI helps the environment: electricity networks, weather, deforestation, agriculture

To discuss: generating a funny meme vs helping to discover a drug — where do we put the cursor?

AI and work

Processed	Endangered Tasks	New professions
Doctor (diagnostic aid)	Repetitive entry	Bias Auditor
Graphic designer (generative tools)	basic translation	AI Ethicist
Prof (custom supports)	Sorting documents	Data Protection Officer (thanks GDPR!)

• History repeats itself: the computer has eliminated jobs... and created much more. But AI will faster → train continuously!

“AI probably won't take your job. But someone who knows how to use it could transform it.”
AI: a tool, not a replacement.

Laws: the European AI Act

• Adopted in 2024 — first major law in the world dedicated to AI
• Gradual application over several years
• The brilliant idea: classify AI by risk level

The European duo:

• GDPR → protects data
• AI Act → regulates the uses of AI

The risk pyramid | Level | Examples | Rule |

|---|---|---|
| Unacceptable | Social rating of citizens, manipulation of the vulnerable | INTERDIT |
| High risk | Recruitment, credit, justice, exams, medical | Strict conditions: tests, transparency, human control |
| Limited | Chatbots, generated images, deepfakes | Must warn: “I am an AI” |
| Minimal | Anti-spam, game AI, recommendations | Free |

Tip: the more AI touches on freedom, money, health, education → the higher it rises in the pyramid.

It’s up to you to classify!

Where do you place these AIs in the pyramid?

1. An AI that suggests a playlist
2. An AI that grades your patent copies
3. A state that awards “good citizen points”
4. A homework help chatbot
5. An AI that decides your student loan

(Answers: 1→, 2→, 3→, 4→, 5→)

The rules of the trial

1. Preparation of teams — 5 min
2. Pleading defense — 2 min flat
3. Plea accusation — 2 min flat
4. Right of reply — 1 min each
5. Jury questions — 3 min
6. Vote (web page, tab 3) + reasoned verdict

Golden rule: we plead the imposed position, even if we think the opposite. It’s strength training for the brain!

The final twist

This case has already happened. In real life.

United Kingdom, 2020: exams canceled (Covid) → an algorithm awards final grades.

Result:

• Underrated disadvantaged high school students
• Demonstrations in the street: “F** the algorithm!”*
• System abandoned in a few days

Does this change your verdict?

What to remember

1. An AI is never fairer than its data (bias)
2. Our data is worth gold; GDPR gives us rights
3. Important decision → require transparency + human control
4. Responsibility is always human, never that of AI
5. AI has an environmental cost — to weigh against the benefits
6. AI: a tool that transforms work, not a replacement
7. The AI Act : a pyramid of risks, from prohibited to free