The Grand Oral Meets AI: Training the Voice Without Outsourcing It

A candidate stands up, alone before the jury, with no notes and no safety net. What the Grand Oral measures is not the tidiness of an outline but the presence of a mind thinking in real time. Yet artificial intelligence, a formidably effective training tool, carries within it the opposite temptation: smoothing the speech until every personal trace is erased.

What the exam actually assesses

The Grand Oral is not a knowledge test in disguise. The official guidelines are clear: it evaluates the quality of sustained, continuous speech, the solidity of the knowledge mobilized, and the ability to argue for a choice — the choice of topic, and more broadly the choice of an intellectual trajectory. Three qualities structure the grading.

• Voice: a controlled pace, eyes that seek out the jury rather than the ceiling, breathing that carries the idea instead of betraying it.
• Spontaneity: the ability to rephrase under the pressure of an unexpected question, without reciting a memorized text.
• Conviction: that almost physical tension that sets apart a candidate who defends an idea from one who merely reports it.

These three qualities share one trait: they cannot be read off a page, they are lived in the moment. That is precisely what training must preserve — and what technological convenience can, with no malicious intent, erode.

What AI does well: structuring, drilling, challenging

Clarifying the architecture of a thought

A student stuck on their outline doesn't need someone to write it for them; they need help seeing where their reasoning limps. A well-used conversational assistant can play that mirror role: asking "why this transition?", flagging a repeated argument, suggesting a more logical progression between parts. The student stays in control of the substance; the tool only lights up the framework.

Training for contradictory debate

This is arguably the most legitimate use. Simulating an objection — "and what if someone answered that…" — forces the student to leave their prepared text behind and improvise a reply. Repeated a dozen times from different angles, this exercise builds exactly what the jury will test: argumentative reactivity, not memorization.

Simulating a demanding jury

Asking follow-up questions, varying the register — sympathetic, skeptical, technical — timing a delivery: AI can recreate pressure close to that of the actual exam, on demand and without social judgment. For a student intimidated by public speaking, this low-stakes repetition defuses part of the performance anxiety that blocks the voice long before the idea runs out.

The risk: a conviction that sounds manufactured

The danger isn't the tool, it's the dependency it quietly installs. A candidate who has an AI draft their entire text, then learns it by heart, produces a speech that is syntactically flawless and rhetorically dead. Grand Oral juries, year after year, report the same symptom: sentences that are too perfect, vocabulary that sounds identical from one candidate to the next, an absence of roughness — that hesitation, that live rephrasing, which signals a mind thinking rather than reciting.

There is an irony worth naming here: the more AI smooths the text, the further it pulls the candidate from what the exam is trying to reveal. A student who knows their subject but stumbles over phrasing is more convincing than one who recites a flawless text whose turns of phrase they never made their own. The jury isn't grading prose; it's assessing a person thinking in front of them.

A concrete method for balance

The dividing line isn't between "with AI" and "without AI," but between building the scaffolding and pouring the concrete for the student. A simple method keeps that boundary in place.

• Phase 1 — Structuring (with AI): the student dictates their ideas in raw form, asks the tool to organize them into an outline, challenges that outline, and reworks it themselves.
• Phase 2 — Phrasing (without AI): the student writes or speaks their own sentences out loud. No generated text is memorized word for word — only the ideas and their sequence are retained.
• Phase 3 — Adversarial drilling (with AI): simulated trick questions, on-the-spot rephrasing, timing.
• Phase 4 — Delivery before a human (without AI): a teacher, a parent, a classmate — a real gaze, the only one able to judge whether the voice carries and the eye contact holds.

This alternation protects the essential: AI prepares the ground, but it is always the candidate's own mouth, under the pressure of the moment, that must produce the final sentence.

A question to carry forward

The Grand Oral has never evaluated a text: it evaluates someone thinking in public. AI can sharpen the architecture of a speech, but it cannot lend its voice without that voice sounding, sooner or later, like someone else's. The question every candidate should ask themselves is therefore not "which tool will save me time?" but "what is the cost, to my own voice, of the time I think I'm saving?"