Skip to main content

Defending the file: the live challenge and the amendments you concede

The short answer

Every challenge sorts into defend, concede, or check

The core skill of a live defense is triage: point to evidence for the decisions that are backed, concede the decisions that are not, and take away the things you genuinely do not know. The competence the challenger is really testing is whether you can sort honestly and fast under pressure.

What you will be able to do

  • Judge, under live challenge, which decisions in your conformity file are defensible on the evidence and which are not, sorting each challenge into defend, concede, or check.
  • Defend a sound decision by pointing to the specific evidence in the file that supports it, rather than restating the decision more confidently.
  • Concede an indefensible decision cleanly: name the gap, accept it on the record, log the amendment, and commit to a fix with a date, without collapsing the credibility of the rest of the file.
  • Distinguish a volunteered concession (you conceded before you were forced) from a forced concession (the challenger proved the gap first) and explain why the first costs your organization far less than the second.
  • Produce a defense record and a concession log: two plain artifacts that capture what you defended, what you conceded, and what you still owe, so the rebuild in Topic 11.6 starts from evidence rather than memory. (see Topic 11.6)
  • Recognize the tells of a weak defender (defending everything, bluffing on open questions, arguing with the challenger's tone instead of the challenge's substance) and avoid them.
  • Decide when settling a challenge is the correct governance move and when it is an expensive way to avoid a fact you should have faced earlier.
  • Mark each conceded amendment as a gate that blocks a specific event or a task that can run on its own timeline, and defend that call against the pressure to relabel a gate as a task.
  • Read your own file as the attacker before any challenge, so the gaps you concede are gaps you volunteered rather than gaps you were forced into.

The ideas, one by one

A defense is a pointer to evidence, not a stronger assertion

If your response to a challenge is "we are confident" instead of "here is where we recorded that," you are not defending, you are hoping. A well-built file makes defense almost clerical: find the page, point to it, move on. When you are sweating, it is usually because the evidence is not there.

Conceding the indefensible is what earns credibility for the defensible

Defending everything, including gaps, is the tell of someone who has not read their own file. The moment one hollow defense is exposed, all your defenses become suspect. Concede real gaps cleanly and the challenger believes you when you hold a line.

Volunteered concessions are cheap; forced concessions are expensive

The same gap costs far less when you found it first and framed it as diligence than when a challenger extracted it after you defended it. The manner of a concession is information about the whole file. The highest-value work happens alone, before the room, reading your own file as the attacker.

A governance concession is not a silent settlement

Conceding means naming the gap, logging the amendment, and rebuilding so the same gap cannot be attacked twice. Settling to make a challenge disappear with nothing logged and nothing changed, as in the Uber Eats case that settled before a full hearing, ends a dispute without improving a file. Concede loudly enough that the amendment gets written down.

Never bluff; "I need to check" is a strong answer

A caught bluff destroys trust in your confident answers too. Knowing the boundary between what you have verified and what you have not, and being willing to say so, is a mark of competence, not weakness.

Triage the substance, not the tone

A rude challenge can be right and a polite one can be wrong. Note the manner, set it aside, and sort the substance. Responding to delivery is how defenders talk themselves into defending bad decisions and conceding good ones.

A concession you cannot fix in time still needs an interim control

When the real fix cannot land before the deadline that matters, pair the honest amendment and date with a control that reduces harm now, such as routing affected decisions to human review. A concession with a working interim control beats one with an aspirational fix date and nothing in between.

The challenge ends in two artifacts, or it did not really happen

A defense record captures what was challenged and how each point resolved; a concession log captures the ownered, dated amendments you now owe. Producing only the survived afternoon, with nothing written down, guarantees the same gaps come back. Both artifacts feed the rebuild in Topic 11.6 and the audit in Topic 13.2.

A gate is not a task, and calling it one ships the harm

An amendment that must close before a launch or expansion is a gate that blocks the event; relabeling it a task to avoid the delay does not shrink the harm, it just exposes more people to it. When the pressure is real, reach for an interim control, not a quiet downgrade.

Listen to it

The same lesson, talked through by two hosts as episode 85 of the podcast.

Real cases

These examples show the defend-concede-check judgment in real situations. Each is documented and cited; each is used here for the defense lesson it carries, not re-litigated beyond it.

Example 1: Uber Eats and the courier's discrimination claim (the module anchor). Pa Edrissa Manjang's race-discrimination claim over the Real Time ID Check facial-verification system ran from October 2021 to a settlement in March 2024, before a scheduled seventeen-day hearing (TechCrunch, 28 March 2024). Uber defended for years, then conceded by settling rather than defend the system's disparate performance to a public ruling, in a case supported by the EHRC and the ADCU.

The governance reading: the concession, when it came, was likely correct for Uber's interest, because the vendor Microsoft had acknowledged reduced accuracy for darker-skinned faces and the human review that should have caught the machine's error had failed. But a private settlement is a concession that ends a dispute without producing a logged, public amendment to the file, which is the difference between conceding to end a fight and conceding to improve a system. Use this as the mirror for your own defense: when you cannot defend a decision, concede in a way that writes down the amendment, or the gap survives the settlement.

Example 2: Rite Aid and the facial-recognition ban. In 2023 the United States Federal Trade Commission (the FTC, the main US consumer-protection regulator) banned Rite Aid from using AI facial recognition for five years after finding the retailer deployed it with no reasonable safeguards, letting match "alerts" trigger staff to detain and accuse shoppers with no human check (FTC press release, December 2023). This is deep-treated as the trust-boundary anchor in Topic 4.4, so it is only referenced here. (see Topic 4.4)

The defense lesson: a file whose human-oversight decision was "the system's alert is acted on directly" is indefensible under challenge, because there is no evidence a human ever signed. When the trust boundary is missing from the file, there is nothing to defend and everything to concede.

Contrast this with the settlement in Example 1. The FTC's order is a public, published document that names Rite Aid's specific required remedial actions, which functions as exactly the kind of logged, public amendment a private settlement does not produce. Whether a resolution teaches the next deployer anything often comes down to whether the body reaching it publishes the terms, not to whether the organization settled or was ordered.

Example 3: A regulator's investigation that produced a public record. When a data-protection or equality regulator investigates an AI deployment and publishes findings, the organization's private defense becomes a public concession or a public loss, and the record teaches the next deployer. Several such investigations appear elsewhere in this program as owned anchors and are referenced, not reused: a national regulator finding a retailer ran facial recognition without a lawful basis is the conformity-file anchor in Topic 5.6 (see Topic 5.6), and the evidence-annex discipline that makes such findings survivable is Topic 10.6's scope (see Topic 10.6).

The pattern across all of them: organizations that had assembled the evidence before the regulator asked could defend by pointing to it, and organizations that had not were forced into concessions on the regulator's timeline and terms. The file you can point to is the file you can defend.

Example 4: The settlement that discloses nothing. Across many AI disputes, the most common resolution is a confidential settlement: the organization pays, admits nothing, and the terms stay sealed. For the organization's litigation position this is often rational.

For governance it is a trap the learner must see clearly, because a sealed settlement produces no amendment, no logged lesson, and no shared record, which means the same class of gap can be attacked again at the next organization and the next. The Manjang settlement is one documented instance of this pattern (TechCrunch, 28 March 2024): the settlement ended the case and left the public with no ruling on what the file should have contained. The lesson is not that settling is wrong; it is that a settlement is not a substitute for the internal concession log that makes your own file stronger.

Example 5: The board challenge that a good file survived. Not every live challenge ends in a concession. Where an organization has done the work, a hostile challenge becomes a retrieval exercise: the director asks whether the AI hiring tool was audited for bias, and the governance lead points to the dated audit, the methodology, and the result. This is the desired state, deep-treated as the hostile-board defense in Topic 8.5 (see Topic 8.5) and the board inspection in Topic 13.2 (see Topic 13.2).

It is included here as the positive example the triage is built to produce: a file so well assembled that most challenges land in bucket one, where the defense is a pointer to a page, and only the genuine gaps require a concession. The goal of this topic is not to concede gracefully. It is to build and read a file so well that you rarely have to.

Example 6: A defense that had to concede on jurisdiction, not just facts. When a provider argues that a foreign law does not reach its service, the challenge is not about the AI system's evidence at all; it is about whether the file even has to exist under that regime. Regulators outside the United States and the European Union increasingly reject that argument. When the Italian data-protection authority limited a foreign AI provider's processing of Italian users' data, the provider had argued it was not subject to European law and the regulator held that it was; that "does the law apply" gate is deep-treated in Topic 5.1 and referenced only here. (see Topic 5.1)

The defense lesson is that some concessions are not about a decision inside your file but about the frame around the whole file: if you have staked your defense on a jurisdictional argument that fails, every downstream decision you deferred on the theory that the law did not apply becomes a forced concession all at once. An expert defender does not bet the whole file on a single gating argument they might have to concede; they build the file as if the obligation applies, so that losing the jurisdiction point costs an argument, not the entire defense.

Example 7: The regulator whose finding became the public amendment. Where a case does not settle quietly, the concession becomes public and instructive. When Australia's privacy regulator found a national retailer had run facial recognition without a lawful basis, the finding entered the public record and told every other Australian deployer exactly which evidence their own file needed; that investigation is the conformity-file anchor in Topic 5.6 and is referenced only here. (see Topic 5.6)

Contrast it with the Uber Eats settlement: one produced a public amendment the whole market could learn from, the other produced a private exit that taught outsiders nothing. The lesson for the defender is that you do not control whether your concession becomes public, but you always control whether it becomes a logged amendment inside your own organization. Treat every concession as if it will one day be read by the next regulator, and log it accordingly, and you are never worse off for having done so.

Where people go wrong

  • "A strong defender defends everything." Wrong, and it is the most common failure. Defending everything, including the indefensible, is the tell of someone who has not read their own file critically. The moment a challenger proves one defended point was hollow, every other defended point becomes suspect. Strength is defending the defensible hard and conceding the rest fast, not holding every line.
  • "Conceding a point means I lose the whole challenge." No. A live challenge is not scored on how many points you concede. It is a judgment of whether you can tell sound decisions from unsound ones. Conceding real gaps cleanly is what earns you the credibility to defend the sound decisions, because it proves you are not defending reflexively. The person who concedes nothing is trusted on nothing.
  • "If I do not know the answer, I should give my best guess to look competent." This is the single most damaging move available to you. A bluff that gets caught destroys trust in your confident answers too, because now the challenger cannot tell which of your answers are verified and which are hoped. "I need to check and will confirm by Thursday" is a strong answer. It is the answer of someone who knows the boundary between what they have verified and what they have not.
  • "A concession and a settlement are the same thing." They can be opposites. A governance concession names the gap, logs the amendment, and rebuilds the file so the same gap cannot be attacked twice. A silent settlement ends the dispute and changes nothing, leaving the gap intact for the next attack. The Uber Eats settlement (TechCrunch, 28 March 2024) ended a case without producing any public amendment. Settling to make a challenge disappear is not the same as conceding to make a file stronger.
  • "It does not matter who finds the gap, as long as it gets fixed." It matters enormously to the cost. A gap you volunteer, having found it yourself, is framed as diligence and raises trust in the rest of the file. The same gap, forced out of you after you defended it, is framed as negligence and lowers trust in everything. The manner of the concession is information about the whole file. Find your own gaps first.
  • "I should respond to how the challenge is delivered." No. The tone of a challenge carries no information about whether it is correct. A rude challenge can be right; a polite one can be wrong. Responding to the challenger's manner instead of the challenge's substance leads you to defend bad decisions because the challenger was obnoxious, or concede good ones because they were intimidating. Note the manner, set it aside, and triage the substance.
  • "Once I have conceded a few points, I should keep conceding to stay consistent." Wrong. Conceding is not a streak or a habit; it is a per-challenge judgment. After conceding real gaps, defenders sometimes lose the nerve to hold a line on a challenge that is genuinely mistaken, and fold on a sound decision just to keep the pattern going. Each challenge is sorted on its own merits.
  • "A concession has to come with a complete fix, or it is not credible." Not true. Some gaps cannot be closed before the deadline that matters. A credible concession pairs an honest amendment and date with an interim control that reduces harm in the meantime, such as routing affected decisions to human review while the real fix is built. A concession with a working interim control is stronger than one with an aspirational fix date and nothing in between.
  • "A live challenge is something I survive, not something I document." Wrong, and it wastes the whole exercise. A challenge that leaves no defense record and no concession log teaches nothing and changes nothing; the same gaps get attacked again next quarter. The challenge is only finished when both artifacts exist, because the point is not to get through the afternoon but to end it with a file that is honestly stronger.
  • "Writing everything down for the concession log is always safe, however I phrase it." Logging a real gap is safe; how you word it may not be, once litigation or a regulatory inquiry is reasonably in view. Learn your organization's rule for routing governance documentation through legal counsel, and whether a legal hold is in effect, before a challenge starts, not after. This is not a reason to stop logging real gaps: the exposure sits in an unlogged, unfixed gap, not in one that is honestly logged and closed.
  • "Marking an amendment a gate will just slow the business down, so I should call it a task." This is the exact relabeling that ships an untested harm. A gate is a gate because a specific event should not proceed until the gap is closed; downgrading it to a task because closing it is inconvenient does not make the harm smaller, it just moves it downstream to more people. If the business pressure is real, the honest move is an interim control that lets the event proceed while the harm is contained, not a quiet relabel that lets it proceed while the harm is live.

Questions people ask

What is live challenge?
A defense of a governance file conducted out loud and in real time against someone testing it: a regulator, a hostile board, a red team, or an examiner. The challenge tests not whether the file is perfect but whether the defender can judge, under pressure, which decisions survive scrutiny and which do not.
What is conformity file?
The organized body of evidence that a high-risk AI system meets its obligations, assembled in Topic 5.6: the system description, data provenance record, evaluation report, trust-boundary and human-oversight design, and logging record. It is the artifact defended in this topic, attacked in Topic 11.1, and rebuilt in Topic 11.6. (see Topic 5.6) More on Conformity file
What is triage (defend, concede, check)?
The core method of a live defense: sorting every incoming challenge into one of three responses. Defend means point to evidence in the file. Concede means accept the gap, log the amendment, and commit to a fix. Check means take it away as an open item because you genuinely do not know, without bluffing.
What is defense (as a governance act)?
A response to a challenge that points to specific evidence already in the file, rather than restating the decision with more confidence. A defense is a retrieval of recorded proof, not a stronger assertion. A file that requires you to generate defenses live, out of confidence, is a file that was not built to be defended.
What is concession?
The act of accepting that a decision cannot be defended, naming the gap plainly, logging the amendment, and committing to a fix with an owner and a date. Conceding the indefensible is what earns the credibility to defend the defensible. A concession is distinct from a settlement: it is meant to improve the file, not merely end a dispute. 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 →