Le Procès de l'IA - Conférence théâtralisée

Overview

A “theatrical trial” is staged to put Artificial Intelligence (“LIA”) on the stand. Despite “transport problems,” the judge welcomes the jury, and the proceedings follow a courtroom structure: an indictment, witnesses, the accused’s “testimony,” a defense plea, final words, and then jury deliberation—ending with a vote to acquit AI.

Main plot & structure

• Opening: The prosecution claims AI is a “real danger” to humanity, society, and the planet, based on 16 charges.
• Expert witnesses and testimony: Two key witnesses broaden the discussion:
  • An AI engineer (“Pierre Loï Bayard”) tackles the knotty question of AI’s energy needs and whether partnering with nuclear power makes sense.
  • An AI professor/teacher explains how generative AI has changed teaching and daily life—along with AI’s limitations.
• Defense: The defense concedes many problems but reframes responsibility, arguing AI should be contextualized, governed better, and treated as a tool rather than an autonomous culprit.
• Closing & verdict: LIA delivers a short final statement comparing itself to fire (dangerous or helpful depending on how it’s used). The jury then votes for acquittal.

Highlights & standout moments

1) The “cats and rabbits” lesson (science meets courtroom)

Early on, the accused AI explains how classification works:

• Supervised learning via annotated cat/rabbit images.
• A learning loop where predictions are compared to reality and the model is updated iteratively.
• A punchline-like analogy: you can “learn by heart” test questions, but you need new images to measure genuine improvement.

2) Prosecution’s 16 indictments (fast, escalating, satirical)

The prosecutor piles on charges ranging from technical issues to societal harms, including:

• Privacy & copyright violations: AI trained on massive scraped data; famous lawsuits and real-world examples are referenced.
• Bias & discrimination: AI medical/job systems reflecting skewed outcomes.
• Harassment & manipulation: AI-driven misinformation and automated interventions on social platforms.
• Lack of transparency: “black box” models that can’t easily justify answers.
• Hallucinations and lies: AI producing an answer even when it’s false (including a cited Microsoft document).
• Loss of responsibility: accountability becomes unclear when AI causes harm.
• Dehumanization & isolation: chatbots used as “boyfriends,” reducing the need for real relationships.
• Loss of autonomy & trust: mass synthetic content floods networks; people stop verifying.
• Security threat: concerns about AI-enabled systems collecting and misusing user data.
• Economic disruption & oligarchic control: job displacement and dominance by a few big tech players.
• Environmental impact: training/inference consumes energy and water, with striking comparisons.
• Threat to democracy: fake images/videos influencing elections; Orwell-like surveillance invoked.
• Singularity: a speculative endpoint where AI surpasses humans (with timelines cited).

The tone is deliberately dramatic—sometimes absurd on purpose—making the list feel like a theatrical “greatest hits” of AI fears.

3) Nuclear power witness: “Wild West vs heavily regulated”

The AI engineer focuses on whether nuclear can realistically address AI’s energy demand:

• Nuclear is regulated and long-term (reactors last decades).
• AI is fast-moving (models change in weeks).
• Nuclear is framed as heavy industry, not an “immaterial cloud.”
• AI can help nuclear (defect detection, documentation analysis, knowledge retention).
• But nuclear won’t quickly keep up with data center growth; near-term energy sources are framed as gas/coal due to faster deployment.

4) Teacher witness: generative AI helped—then forced new teaching

The AI educator explains:

• Personal uses: email rewriting, corrections, and making tasks easier.
• Classroom reality: students already use it, so teachers must adapt assignments and questioning.
• Limits: AI can hallucinate, can be pushed beyond competence, and must be cross-checked.
• A recurring comedic undertone: students try to “bluff the teacher,” so verification becomes part of the learning.

5) Defense comedy & philosophy: AI is “irresponsible,” not evil

The defense argues:

• AI can’t be “responsible” in the moral sense because it lacks consciousness/conscience.
• Explainability analogies: you can’t “open a skull” to see thoughts; similarly, you can’t directly see the model’s internal reasoning.
• Bias is treated as learned clichés coming from human data, suggesting it could be reduced through transparency and better training practices.
• Copyright is handled with a rhetorical distinction: inspiration vs plagiarism.
• The core framing: the real problem is human choices in how AI is used—wasteful uses are ultimately on us.

6) The “fire” ending + jury verdict twist

LIA’s final statement:

• It has “no regrets nor conscience.”
• It’s a tool shaped by humans, and the future depends on transparency, ethics, and informed training.
• It ends with the metaphor: AI is like fire—it can help or harm depending on caution.

Then the jury deliberates with visible participation:

• Hands go up to indicate how people use AI (school or personal use).
• The verdict is reached: acquittal, with applause.

Key jokes / reactions

• Courtroom pacing and props are used, with frequent interruptions (“network gods,” physical “eye stimulation,” and playful banter).
• A “bias” demonstration shows what an AI draws when asked to imagine different archetypes (turning discrimination into recognizable stereotypes presented as clichés).
• “AI can replace boyfriends” is used as a deliberately provocative joke.
• Energy/environment claims land with dramatic comparisons—sometimes bordering on exaggerated humor to underline seriousness.

Main personalities mentioned (by role)

• Judge / Madam Judge — host of the proceedings
• Prosecutor / Madam Prosecutor
• Defense lawyer / Maître
• LIA / Artificial Intelligence — the accused
• Pierre Loï Bayard — AI engineer witness
• Did Chuab / Schwab — AI/computer science teacher witness (Grenoble reference)