Skip to main content

The viva: defending your command live against the examiner

The short answer

The viva tests command, not knowledge

Recall was tested in writing. The viva tests whether you can defend the specific decisions in your own dossier against an examiner who has read all of it. Studying frameworks for a viva is preparing for the wrong exam; study your own record until you know exactly where your evidence ends.

What you will be able to do

  • Evaluate a live examiner's challenge to one of your command decisions and judge whether the decision is defensible on the evidence you actually have, indefensible and requiring concession, or somewhere in between requiring a qualified answer.
  • Distinguish a defensible answer (grounded in a dated, inspectable artifact in your dossier) from a bluff (an assertion the record does not support), and explain why the second one destroys your credibility on everything else.
  • Trace any answer you give back to the specific artifact that backs it: the AI systems inventory, the conformity file, the evaluation report, the incident log, the evidence annex. (see Topic 13.1)
  • Concede an indefensible decision cleanly, naming the gap, the risk it created, and the concrete remediation, without collapsing your standing on the decisions that are sound.
  • Judge the difference between an examiner attacking your evidence and an examiner attacking you, and respond to each without losing composure or ceding a defensible position.
  • Construct, on the spot, a defense of a decision the examiner frames in a way you did not anticipate, by returning to the underlying principle and the record rather than reaching for a rehearsed script.
  • Assess your own viva performance against the standard the AIGP (Artificial Intelligence Governance Professional) credential tests for and then goes past: not "do you know the rule" but "did you apply it, and can you defend the application."
  • Recognize the seven recurring challenge archetypes an examiner uses (show-me, edge case, gap probe, personal, counterfactual, self-location, temporal drift) and reach for the correct response to each even when the specific question is one you did not prepare.
  • Deliver a viva answer in the four-beat structure (name the decision, name the evidence, state the tradeoff, hold or concede) automatically enough that you never freeze and never wander outside your record under pressure.

The ideas, one by one

The boundary principle governs everything

You can defend what your documented record supports and nothing else. The examiner compares your spoken answers to your written commitments in real time, so the moment your voice claims more than your record backs, the gap becomes the story. Preparation is calibration to your own evidence, not rehearsal of eloquent answers.

Every answer is one of three kinds

A full defense (grounded in a dated artifact, held and stopped), a qualified answer (sound reasoning, partial evidence, gap named before the examiner names it), or a concession (indefensible, conceded cleanly with a remediation). The skill is knowing which kind you are in within seconds of the challenge landing.

A clean concession is a display of command, not weakness

Conceding the genuinely indefensible while defending the sound is exactly what the viva measures. Defending an indefensible decision does not save it; it destroys your credibility on every defensible one, because you have shown your judgment cannot tell a good call from a bad one.

Answer in four beats

Name the decision, name the evidence, state the tradeoff, then hold or concede. If you cannot complete the second beat, you are already outside your record and should go straight to an honest concession rather than bluffing.

Distinguish the evidence attack from the personal attack

Meet a challenge to your record with evidence or a clean concession. Meet a challenge to you with a calm route-back to an artifact, never with a defense of your ego and never with overreach. The build-before-you-govern spine exists so you can answer the personal attack by pointing at the model you built and broke.

Intent is not evidence

"We were going to run that audit" defends nothing. You can defend what you did, not what you meant to do; meaning to do the work lives in concession territory. Say it, back it with the artifact, stop.

A thorough dossier is necessary but not sufficient

The viva tests whether you understand the record, not whether it exists. Someone who inherited a complete file they did not build fails on the first follow-up. The dossier is the ground; command of it is what the viva measures.

The bluff count is the number that matters

In practice and in reality, a viva with zero bluffs and several clean concessions is a strong pass; a viva with one undetected bluff is a fail, because in a real proceeding that single bluff is where you lose everything. Find your bluffs in a simulated viva so a real examiner never finds them.

This is the closest simulation of what actually happens to AI leaders

The regulator interview, the deposition, the board's private session, the acquirer's diligence, all are vivas by another name, and all are governed by the boundary principle. The graduate who can sit this viva has crossed a line the AIGP exam does not draw. (see Topic 13.4)

Challenges come in a small number of shapes

Show-me, edge case, gap probe, personal, counterfactual, self-location, and temporal drift cover nearly everything an examiner asks. You cannot rehearse every question, but you can rehearse the seven responses, so an unfamiliar question resolves into a familiar move.

The language of command points at inspectable artifacts

The difference between a bluff and a defense is often the exact words: the weak answer leans on confidence, habit, or ego, and the strong answer names a specific dated artifact and stays inside what it supports. Make every clause point at something the examiner can check.

Silence is often a test, not a verdict

After a held answer, an examiner may go quiet to see whether you will keep talking and drift outside your record. Holding the silence is the disciplined move; filling it usually hands the examiner a new opening.

Candor is the only rational strategy under a fixed record

When your decisions are documented, hiding a gap or overstating a control does not reduce your exposure; it compounds it, because the misstatement is now a second, worse problem layered on the first. The Kramer robocall reckoning shows it plainly: a fixed record made candor the only move that did not make things worse.

The artifacts you wrote honestly are the ground you stand on

Every earlier module's artifact becomes part of the boundary you defend in the viva. Honest disclosure letters, evaluation reports, and decision memos give you defensible ground; ones written to look good give way exactly where an examiner pushes. The viva is where the honesty of your earlier work is finally priced.

Being examined well teaches you to examine well

The graduate held to the boundary principle becomes the person who demands it of vendors, teams, and systems they approve. That is how the discipline raises its own standard: rigorous, fair examination propagates through the people who survived it.

Listen to it

The same lesson, talked through by two hosts as episode 97 of the podcast.

Real cases

These are real accountability moments where a person, or an institution, had to defend AI-related decisions against an examiner who had studied the record. The lesson each carries is the boundary principle: the record set what could be defended.

Example 1: The New Hampshire robocall reckoning (the anchor). After the January 2024 deepfake Biden robocalls, the accountability came in layers, and each layer worked from a fixed record. The FCC proposed a 6 million US dollar penalty against Steve Kramer and finalized it in 2024, having established that he drafted the script, commissioned the AI voice clone, chose the spoofed number, and directed the calls (Federal Communications Commission, FCC 24-59, Notice of Apparent Liability, 2024; Forfeiture Order, 2024). Separately, Lingo Telecom, the carrier that transmitted the calls, faced a proposed 2 million US dollar penalty and settled with the FCC for 1 million US dollars plus a binding compliance plan requiring strict adherence to the STIR/SHAKEN caller-ID authentication rules (Federal Communications Commission, "FCC Settles Spoofed AI-Generated Robocalls Case," 21 August 2024). New Hampshire indicted Kramer on thirteen felony voter-suppression counts and thirteen misdemeanor candidate-impersonation counts, though a state jury acquitted him on those charges in June 2025, accepting his argument that the statutes did not reach an unsanctioned primary rather than disputing what he had done (New Hampshire Department of Justice, 2024; New Hampshire v. Kramer, jury verdict, June 2025). The teaching point is not the size of the penalties, nor even the criminal outcome. It is that in every forum, the decisions had been documented before the defense began, and the documentation, not the later explanation, set the boundary of what anyone could argue. Even in the forum where Kramer prevailed, he did so by contesting the law, not the record of his own conduct, which he did not deny. The carrier's defense was narrow and led to a settlement precisely because the record showed it had failed a specific, nameable control. Command, or its absence, was legible in the record.

Example 2: A congressional hearing as a viva. When AI executives are called before legislative committees, the format is a viva in all but name. The members have staff-prepared records, they ask a decision to be defended, and the answer is measured against the public record of what the company did. The executives who fare worst are those who assert capabilities or safeguards the record does not support and are then confronted with the contradiction. The pattern is consistent across many such hearings and is worth studying as a genre: the boundary principle governs a legislative hearing exactly as it governs a doctoral defense. Claims that exceed the documented record become the story. (Illustrative of the format generally; the specific dynamics repeat across multiple US and EU hearings from 2023 onward.)

Example 3: The litigation deposition. In a deposition, opposing counsel has your documents and questions you under oath about the decisions in them. It is the purest evidence-attack viva that exists. The Uber Eats courier case, where a Black driver's repeated facial-verification failures led the company to settle rather than defend the discrimination claim to a full hearing, turned on exactly this dynamic: once the record of repeated failures was established, the defensible territory shrank to the point where settlement was the rational move. (see Topic 11.2) The deposition does not care about your eloquence. It cares about what your documents show and whether your sworn answers stay inside them.

Example 4: The regulator interview. European data-protection authorities and sector regulators increasingly conduct interviews as part of AI investigations, sitting the responsible officer down and asking them to defend specific processing decisions against the file. The Italian Garante's action against the maker of the Replika chatbot, which banned the service for failing to protect minors and later imposed a 5 million euro fine, developed through exactly this kind of record-based examination of what the company knew and decided. (see Topic 11.6) The officer who can point to a dated impact assessment defends; the officer who cannot, concedes. The interview is a viva conducted by someone with statutory power.

Example 5: The acquirer's due diligence. When one company acquires another, the acquirer's technical and legal teams grill the target's leaders on their AI systems: what is in the training data, what is the provenance, what are the open liabilities. This commercial viva is where an AI systems inventory that was assembled honestly pays off and one that was assembled to look good falls apart. (see Topic 8.3) The diligence team has read the data room; the target's officer defends against that record. A gap discovered here does not fine you; it discounts the price or kills the deal. The stakes teach the same lesson: build the record honestly, because someone who has read it will one day ask you to defend it out loud.

Example 6: The internal board's private session. The most frequent viva an AI leader faces is the least visible: the board's private session, where directors who have read the pre-read ask the officer to defend a decision without management present to help. This is where the dossier assembled in Topic 13.1 either carries the officer or exposes them. (see Topic 13.1) The board is not hostile, but it is examining, and the officer who understands their own record defends it while the officer who inherited it flounders on the first follow-up.

Example 7: The whistleblower's own viva. Accountability runs in both directions. A person who raises an alarm about their own organization's AI is also examined against the record, by the organization, by regulators, and sometimes by the public. When a former OpenAI researcher publicly argued the company's training practices violated copyright and offered to provide documents, the strength of the position rested on what could be shown, not on the sincerity of the concern. (see Topic 11.5) The lesson generalizes: whether you are defending your command or challenging someone else's, the boundary principle governs. Claims are worth exactly what the record behind them can support. (This case is referenced with care; the individual was later found dead, and the point here is strictly about record-based accountability.)

Example 8: The court that demanded the record be inspectable. In State v. Loomis, the Wisconsin Supreme Court allowed a risk-assessment tool at sentencing only on the condition that its opacity and limits be disclosed in writing, so that the decision could be inspected. (see Topic 13.1) The court was, in effect, insisting on a defensible record before it would let an AI-informed decision stand. That is the institutional form of the same demand a viva makes of a person: a decision is only as good as the inspectable reasoning behind it, and a decision no one can examine is a decision no one should trust.

Example 9: The hallucinated-citations sanction as a failed viva. When a federal judge fined two lawyers whose brief carried dozens of AI-invented citations, the courtroom exchange that preceded the sanction was a viva in miniature. The judge asked the lawyers to defend the filing, and the record, the fake cases, could not be defended because the cited authorities did not exist. (see Topic 11.1) The lesson is the boundary principle at its harshest. There was no artifact behind the claims, so there was nothing to defend, and no amount of eloquence could conjure a record that was never there. A viva cannot be won by confidence when the underlying evidence is fabricated; it can only be lost more slowly.

Where people go wrong

  • "A viva is a knowledge test, so I should study the frameworks harder." Wrong target. The viva does not test whether you know the EU AI Act or the NIST AI RMF; the written assessments already did that. It tests whether you can defend the decisions in your own dossier. Studying frameworks the night before a viva is preparing for the wrong exam. Study your own record until you know exactly where your evidence stops.
  • "I should defend every decision; conceding makes me look weak." This is the most damaging misconception in the topic. Defending an indefensible decision does not protect that decision; it destroys your credibility on every defensible one, because you have shown the examiner your judgment cannot tell a good call from a bad one. A clean concession with a named remediation is a display of command. The person who concedes nothing is trusted on nothing.
  • "If I sound confident enough, I can carry an answer past a weak point." Confidence is not the currency. The examiner is comparing your words to your record, not weighing your tone. An eloquent claim that the record does not support is a bluff, and a good examiner detects the gap regardless of how you sound. Calibration beats confidence every time.
  • "The examiner attacking me personally means I am losing." Not necessarily. A personal attack is often a probe of your composure, not a substantive point against your record. Treat it as a test to route back to the work, not as a wound to defend. Losing composure over a personal attack, or overreaching to prove yourself, does more damage than the attack itself.
  • "I can defend a decision by explaining what I intended to do." Intent is not evidence. "We were going to run that audit" defends nothing; the audit either exists in the record or it does not. The boundary principle is unforgiving here: you can defend what you did, not what you meant to do. Meaning to do the work lives outside your record, in concession territory.
  • "A longer, more detailed answer is a stronger answer." Padding is a tell. A held answer names the decision, the evidence, and the tradeoff, and then stops. Candidates who keep talking after they have made their point usually do so because they are anxious, and the extra words often wander outside the record and hand the examiner a new opening. Say it, back it, stop.
  • "My dossier is thorough, so the viva is a formality." A thorough dossier is necessary but not sufficient. The viva tests whether you understand the dossier, not whether it exists. An officer who inherited a complete, well-organized file they did not build will fail on the first follow-up question, because thoroughness of documentation and command of it are different things. (see Topic 13.1)
  • "I can defend a decision by pointing to what the industry usually does." Wrong. Industry practice is not your record. "Everyone uses that vendor" defends nothing about whether you assessed the vendor. The examiner is testing your decision and your evidence, not the market's habits. Common practice may inform a decision, but the artifact that records your own assessment is what you defend.
  • "Silence from the examiner means my answer failed." Not necessarily. A skilled examiner often stays silent after a held answer to see whether you will keep talking and wander outside your record. Silence is frequently a test of your discipline, not a verdict on your answer. Fill it and you may hand the examiner a new opening; hold it and you pass the test the silence was.
  • "A counterfactual question is a trap I should refuse." A counterfactual ("why not automate this?") is not a trap; it is a fair test of whether you understood your tradeoff. Refusing to engage it, or claiming there was no alternative, signals you backed into the decision. Name the alternative you rejected and why, and the counterfactual becomes a chance to show command.
  • "If I concede one decision, I have failed the viva." The opposite. Conceding the genuinely indefensible while defending the sound is exactly the calibration the viva measures. A viva where you conceded two decisions cleanly and defended nine on the record is a strong pass. A viva where you defended all eleven, including the two that were indefensible, is a fail disguised as a win, and a real examiner sees through it.

Questions people ask

What is viva voce (viva)?
From the Latin for "living voice," an oral examination in which the person who did the work defends it face to face against an examiner who has already studied the record. In this program, the live defense of your AI governance command against an AI examiner that has ingested your entire dossier.
What is the boundary principle?
The rule that you can defend only what your documented record supports, and nothing outside it. Because the examiner compares your spoken answers to your written record in real time, any claim that exceeds the record creates a contradiction the examiner will find. Preparation for a viva is calibration to this boundary, not rehearsal of answers.
What is full defense?
A viva answer to a challenge that lands inside your evidence. It names the decision, names the dated artifact that backs it, states the tradeoff accepted, and stops. The strongest of the three kinds of answer, available only when the record supports the claim.
What is qualified answer?
A viva answer to a challenge where the reasoning is sound but the evidence is partial. The move is to defend the reasoning while naming the evidence gap out loud, before the examiner names it, keeping the answer inside the bluff boundary.
What is concession?
A viva answer to a challenge that lands outside your evidence, where the decision was wrong, the control was untested, or the artifact does not exist. A clean concession admits the gap plainly, names the risk it created, and names a concrete remediation. Conceding the genuinely indefensible is a display of command, not weakness. More on Concession

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →