Telling the customer: the disclosure decision and the words you actually send
The short answer
Disclosure is a decision, not a reflex
It has four dimensions you reason through and defend: whether to tell, whom, how much, and by when. Dumping everything in a panic and saying nothing and hoping are both the absence of a decision. A defensible disclosure decision, including a defended decision not to tell some group, is the artifact this topic produces alongside the letter.
What you will be able to do
- Analyze an AI incident to decide what must be disclosed, to whom, and by when, separating the legal floor (what a law requires) from the trust floor (what an honest organization owes even when no law compels it).
- Distinguish genuine disclosure from "quiet" non-disclosure: information that is technically available but placed where the affected person will not find it (the CNET hover-over byline) is not disclosure, and treating it as such is the classic failure.
- Identify the triggers that raise a disclosure duty (ongoing harm, affected people who must act to protect themselves, a legal or contractual notification rule, a public discovery about to happen anyway) and weigh them against the reasons teams stay silent.
- Compose the words you would actually send: a disclosure letter that states plainly what happened, what you got wrong, who is affected, what you have done, what the reader should do, and how to reach a human, with accountability in the first person.
- Detect and rewrite the wording patterns that turn a disclosure into a liability: passive voice that hides the actor, blaming the model ("the AI hallucinated"), euphemism that minimizes, over-lawyered fog, and false reassurance.
- Decide the timing: when to self-disclose ahead of discovery, when a correction or editor's note is the right instrument, and why being caught concealing is almost always worse than the original error.
- Calibrate the apology and the remedy: offer a specific apology for the specific harm, and pair the words with a real make-good for anyone who acted on the failure, since a disclosure cannot recall harm already done.
- Defend the disclosure decision as a record, so that when the file is later attacked or a regulator asks, the choice reads as a reasoned judgment made on principle, not a scramble.
The ideas, one by one
Available is not disclosed
A fact the affected person will not encounter in the normal course, the hover-over byline, the footer, the terms clause, is a technicality, not a disclosure. CNET's "quietly, not in secret" is the canonical failure: from the reader's seat, a fact you cannot see is a fact you were not told. Test every disclosure by whether the person actually learns what they need, where they would see it.
Own it in the first person; never blame the tool
"The AI hallucinated" is grammatically true and read by every customer as a refusal of responsibility, because the AI did not choose to deploy itself. The organization is the actor. First-person accountability ("we shipped it, we did not catch it") is not weaker; it is the only version a reader believes, and it protects your individual people while the organization owns the outcome.
The legal floor is not the ceiling
A notification law (like the seventy-two-hour breach duty under the European Union's General Data Protection Regulation) is a mandatory floor, but many AI failures trigger no such law and still demand disclosure on the trust floor. Verify the legal duty with counsel; then ask separately what honesty requires. Organizations people trust disclose on the trust floor, not just the legal one (see Topic 5.7).
The letter has a stable anatomy
Six parts: what happened (plain), what you got wrong (specific), who is affected and how they can tell, what you have done, what they should do, and how to reach a human. A strong disclosure can account for each, including a reasoned omission. Underneath sits one tone: plain, direct, first-person, free of euphemism and fog.
Six wording patterns turn a disclosure into a liability
Passive voice that hides the actor, blaming the tool, euphemism and minimization, over-lawyering, false reassurance, and false urgency (manufactured drama that reads as theater when the letter's own facts do not support it). The analytic core of the topic is reading your own draft as the affected person and rewriting every failing sentence toward the plain, first-person, specific truth.
Self-disclose before you are caught
The prompt disclosure of an honest partial account beats the delayed disclosure of a complete one, because the delay reads as concealment. Move when the post-incident review gives you a defensible account; a statement timed to a journalist's call is exposure managed, not honesty offered, and readers can tell.
Match the instrument to the reach
A direct notice to identified users, a correction on affected content, a public statement for a broad audience, a regulator filing where a duty attaches, often more than one. One statement does not inform everyone; map each affected group to the instrument that actually reaches it.
A disclosure cannot do the fix's job
It cannot recall the harm already acted on and it cannot substitute for a real remediation. Its credibility rests on the "what we have done" being true, and for anyone who acted on the failure, words without a make-good are a disclosure attached to an unfixed problem. Sometimes the discipline is to say less, on principle, when the omission protects the affected person rather than you.
Apologize for the specific harm, once, plainly
A cold no-apology reads as an organization protecting itself; a lavish "we regret any inconvenience" apologizes for a feeling instead of the real harm and hedges the harm away. Name what you are sorry for and to whom, offer it once without drama, and let your actions carry the rest. Saying sorry for a specific harm is a moral act, not only a tactic.
A business-customer disclosure is not a consumer disclosure
A contract often sets stricter notification terms than the law, your business customer may have to disclose downstream to the people actually affected, and the instrument and detail differ. But the anti-patterns do not: no tool-blaming, no fog, and the trust floor still sits atop the contractual one.
You can disclose while the incident is still live
When harm is ongoing, do not wait for a complete review: state what you know, label what you do not, commit to an update cadence, and never speculate as if it were fact. The prompt, honest, partial disclosure protects people during the window the complete account would have missed.
This artifact closes the shipping chain and will be tested for honesty
The disclosure letter completes shipped feature, incident log, post-incident review, disclosure. A Module 11 defense of the file will test whether your disclosure was honest or a cover (see Topic 11.2), and the incident class it describes is what your Module 4 evaluation suite must be built to catch (see Topic 4.2).
Listen to it
The same lesson, talked through by two hosts as episode 23 of the podcast.
Real cases
These examples show disclosure decisions and disclosure wording in real cases, with the reasoning made explicit. The deep anchor is the CNET case; the others sharpen a specific point and are treated in depth by their owner topics.
Example 1 (the anchor): CNET's finance explainers and the word "quietly." Beginning November 2022, CNET (owned by Red Ventures) published seventy-seven personal-finance explainers written by an internally built AI tool, under the byline "CNET Money Staff," with AI authorship visible only on hovering over the byline. After Futurism reported the practice and then surfaced factual errors, CNET's own review found errors in forty-one of the seventy-seven articles, including a compound-interest explainer claiming a $10,000 deposit at three percent would yield $10,300 in a year (the correct figure is about $300), annual-percentage-rate versus annual-percentage-yield confusions, and near-verbatim copying from human writers. The editor in chief, Connie Guglielmo, published a note to readers in January 2023, paused the AI tool, appended corrections and editor's notes, and changed the byline to a clearer disclosure, while reportedly framing the earlier approach in a staff meeting as done "quietly," not "in secret" (The Verge, Futurism, and CNN Business, January 2023). Read against this topic: the whether/whom/how-much/by-when decision was made badly on every axis; the "quiet" hover-over byline is the canonical case of information available but not disclosed (3C); and the defensive framing ("quietly") is the euphemism pattern (3G) that cost more trust than the errors. The lasting consequence underlines the stakes: Wikipedia's editors downgraded CNET from a "generally reliable" source after the episode (Futurism, 2023). A disclosure decision made deliberately and early, in plain first-person words, was available the whole time, and its absence, not the model's errors alone, is what defined the case.
Example 2 (a disclosure done reasonably, for contrast): the Cursor support-bot fabrication. When an AI support agent at the developer-tools company Cursor invented a nonexistent account policy and users canceled in response, a co-founder posted publicly, confirmed plainly that no such policy existed, attributed the failure to the AI system in a way that still owned the company's responsibility, and the company moved to label AI-generated support replies (2025). The deep treatment of that incident belongs to its owner topic (see Topic 1.6); here it is the contrast to CNET on wording and timing. The disclosure was reasonably prompt, plainly worded, and did not hide behind a hover-over technicality, which is why it read as a company catching its own failure rather than one exposed. It is not a perfect model (attributing to "the AI" edges toward the tool-blaming pattern of 3E, and a stronger version would center the human decisions more), but it shows the difference that plain, timely, owned wording makes to how a disclosure lands.
Example 3 (non-disclosure as the failure): fabricated AI author profiles. A publisher discovered to have run articles under fabricated author identities supplied by a third party, with no clear disclosure that the content was AI-generated, faced the same core failure as CNET on the disclosure axis: the affected people (readers) were never plainly told what they were reading. The deep case is owned elsewhere (see Topic 0.3); it is cited here because it is the same lesson from a different industry: the absence of a plain, findable disclosure, more than the use of AI itself, is what breaks reader trust. Disclosure is not a courtesy layered on top of an AI product; it is part of the product's honesty.
Example 4 (the passive-voice trap, illustrative method): "errors were identified and corrected." Consider the common corporate line issued after an AI incident: "Errors were identified and have been corrected, and we are reviewing our processes." Every word is true and the sentence discloses almost nothing: it names no actor (who made the errors?), no substance (errors in what?), no affected party (did this touch me?), and no protective action. This is not a real example from a single company because it is nearly universal; it is the passive-voice pattern of 3G in its natural habitat. The analytic skill is to read such a sentence and see the four things it hides, then rewrite it into the six-part anatomy of 3F. A disclosure that could have been issued by any company about any incident has disclosed nothing specific to yours.
Example 5 (the legal floor is not the ceiling): a data-breach notice versus an AI-error disclosure. When an AI system's failure also exposes personal data, a legal notification duty may attach (for example, the seventy-two-hour authority-notification and high-risk individual-notification duties under the European Union's General Data Protection Regulation), and meeting that duty is mandatory. But many AI failures, like CNET's wrong financial figures, expose no personal data and trigger no breach-notification law, and a team that treats "no legal duty" as "no disclosure duty" has confused the floor for the ceiling. The floor itself is not static: from 2 August 2026, when the European Union AI Act's transparency obligations take effect, a CNET-shaped case (AI-written text on a public-interest topic, published without genuine human review) would separately trigger the AI Act's Article 50 transparency duty, a legal floor that did not exist for the 2023 case. The regulatory-response discipline is owned by a later topic (see Topic 5.7); the point here is that the trust duty (3D, trigger 4) often requires disclosure the law does not, and even where a new law now catches yesterday's trust-floor case, the organizations people trust are the ones that would have disclosed anyway. Verify whether your specific incident triggers a legal duty; then ask, separately, what honesty requires regardless.
Example 6 (owning it in the first person, applied to your own estate). Suppose an AI feature your organization ships produced a wrong output that reached customers. Two disclosures are available. The first: "An automated system generated inaccurate results, which have been addressed." The second: "We deployed a tool that gave some of you wrong information, we did not catch it before it reached you, and here is what happened, whether it affected you, and what we are doing." The first blames the tool and hides the actor; the second owns the outcome in the first person. The customers' trust responds to the second, because it is the only one that treats them as people owed the truth rather than a public-relations problem to be managed. This is not a hypothetical unique to any company; it is the choice available in every AI incident disclosure, and it is the choice your artifact makes in advance.
Example 7 (the imminent-discovery calculus, illustrative). Consider a team that discovers, through its own monitoring (see Topic 3.4), that a feature has been producing a subtly wrong output for a week, and that no outsider has yet noticed. The calculation is not "can we avoid disclosing" but "self-disclose now, or wait and risk exposure." The triggers of 3D favor prompt self-disclosure: there may be ongoing harm, and discovery, while not certain, is plausible. The team that self-discloses converts a potential scandal into evidence of a functioning governance process; the team that waits bets its reputation on nobody looking, which is the CNET bet, and CNET lost it. The lesson is method: when you find your own failure before the world does, that head start is an asset to spend on integrity, not a window to hope it stays hidden.
Example 8 (the correction as an instrument, and its limits): the appended editor's note. When a failure lives in published content, the natural instrument is a correction or editor's note appended to the affected item, because the affected people are its readers and that is where they encounter it. CNET's appended corrections and editor's notes were the right instrument; their weakness was timing (after exposure) and prominence (quiet). The general lesson is that an editor's note works only if it is findable and honest: a correction buried at the foot of a long article, or one that says "this article has been updated" without saying what was wrong, repeats the "available, not disclosed" failure inside the very instrument meant to disclose. A good correction states plainly what was wrong, when, and that an AI tool wrote it, at the top where the reader will see it.
Example 9 (the ongoing-incident disclosure, illustrative method): the live status update. Consider an organization whose AI feature is actively producing a harmful output right now, before the incident is understood. Waiting for a complete account would leave people exposed during the very window they could act. The method (from 3M) is a live status disclosure: state what is known ("since [date] our tool has been showing some users an incorrect balance"), label what is not ("we are still confirming how many accounts and which dates"), commit to a cadence ("we will update by [time]"), and avoid speculation dressed as fact. This is not a single company's case because it is a recurring shape; it is the discipline that lets honesty and speed coexist when the review is not yet done, and it is the antidote to the "wait for the full picture" delay that reads as concealment.
Example 10 (the apology that fails by hedging, illustrative). Consider the near-universal corporate line "we sincerely apologize for any inconvenience this may have caused." Read it as the affected person: it apologizes for "inconvenience" (a mild annoyance) when the harm may have been a wrong medical or financial figure they relied on, and the hedges "any" and "may have" quietly suggest the harm might not have been real at all. The sentence performs contrition while denying the substance of the wrong. The fix is a specific apology tied to the specific harm ("we are sorry we gave you a wrong figure that you may have used to make a decision"), which cannot be read as either evasion or theater. The lesson is method, not a single company: an apology that names the real harm is believed; one that apologizes for a feeling is heard as a dodge.
Where people go wrong
- "We made the information available, so we disclosed it." Available is not disclosed. A fact the affected person will not encounter in the normal course, the hover-over byline, the terms-of-service clause, the footer, is a technicality you are holding in reserve, not a disclosure. The test is whether the person actually learned what they needed to know, in a place they would see it. CNET's "quietly, not in secret" is this mistake said out loud.
- "The AI hallucinated, so we should say the AI made a mistake." The AI did not choose to deploy itself to real users without adequate review; you did. Blaming the tool is grammatically true and read by every customer as a refusal to take responsibility. Own it in the first person: "we shipped a tool that produced false statements and we did not catch them."
- "A vague statement protects us legally." Vague statements protect you until the specifics come out, at which point the vagueness reads as concealment and costs more than the plain truth would have. Involve counsel to get the facts and commitments right, then say them plainly. Accuracy is defensible; fog that looks like a cover-up is not.
- "If no law requires disclosure, we do not have to disclose." The legal duty is a floor, not a ceiling. Many AI failures (like wrong financial figures that expose no personal data) trigger no notification law and still demand disclosure on the trust floor. Confusing "no legal duty" with "no disclosure duty" is how organizations that merely comply lose the trust that ones people rely on keep.
- "We will disclose once we have the full picture." The prompt disclosure of an honest partial account beats the delayed disclosure of a complete one, because the delay is what reads as concealment. Move when the post-incident review gives you a defensible account of what happened and who is affected; follow with detail as you confirm it. Waiting for a perfect, reassuring version means waiting past the moment integrity was possible.
- "'We are committed to transparency' shows we take it seriously." Asserting your own trustworthiness in the same letter where you explain that you failed the reader is empty and often counterproductive. You show seriousness through the specific actions in the "what we have done" section, not through adjectives about yourself. Cut the self-praise; keep the facts.
- "Disclosure means confessing everything we know." Over-disclosure is a real failure: dumping raw internal detail, naming an individual employee, speculating about unconfirmed causes, or exposing another person's data. The standard is "what does this person need to know and do," not "everything." Saying less, deliberately, for the reader's benefit is discipline; it is the opposite of concealment only when the omission protects the reader, not you.
- "Mistakes were made is a fine way to put it." Passive voice deletes the actor, which is exactly why teams reach for it and exactly why readers distrust it. Name who did it: "we made mistakes." A disclosure with no human subject who owns the outcome has not owned the outcome.
- "We will say this will never happen again." You do not know that, and if it recurs the promise becomes the story. Commit to what you control: the tool is paused, the process is changed, and you will not restart until the review catches this error class. A promise you can keep rebuilds trust; a promise you cannot keep spends it twice.
- "One public statement informs everyone affected." Different affected groups need different instruments: a direct notice to identified users, a correction on affected content, a public statement for a broad audience, a regulator filing where a duty attaches. Issuing one statement and treating the whole population as informed leaves the people who most needed a direct word uninformed.
- "The letter is the response." A beautifully worded disclosure attached to a harm you have not actually fixed is a promise you are already breaking, and for people who acted on the failure the words are not a remedy. Disclosure is the public face of a real remediation, including a make-good where one is owed, not a substitute for it.
- "Disclosing through our AI chatbot or a no-reply address is fine." An AI incident disclosed through an automated channel that cannot answer a follow-up repeats the original failure. Give affected people a real way to reach an accountable human; the channel is part of the disclosure's honesty.
- "Getting ahead of the story by issuing a statement the day the reporter calls counts as self-disclosure." Readers can usually tell the difference between honesty offered and exposure managed. A statement timed to a journalist's call is not self-disclosure; it is being caught, dressed as candor. The head start you get from finding your own failure first is only worth something if you spend it before the world forces your hand.
- "We should not apologize; sorry is an admission." A cold, apology-free letter reads as an organization shielding itself, and a specific apology for a specific harm is usually the stronger position. Counsel can advise wording for your jurisdiction (in some places an expression of regret is legally distinct from an admission of fault), but stripping all regret out produces a letter no one believes and no one forgives.
- "A contract's notification clause is the whole of our duty to a business customer." The contract is a floor, often stricter and faster than the law, but it is still a floor. Your business customer may need to disclose downstream to the people actually harmed, so "how much" includes what they need to pass on, and the trust floor sits atop the contractual one.
- "We cannot disclose until the incident is fully understood." When harm is ongoing, you disclose what you know now, label what you do not, and commit to updates, rather than leaving people exposed during the window they could act. Waiting for the full picture is exactly the delay that reads as concealment if the failure surfaces first.
Questions people ask
- What is disclosure decision?
- The deliberate judgment, made on four axes (whether to tell, whom, how much, and by when) and two bases (the legal floor and the trust floor), about what to communicate after an AI incident. It is one of the two artifacts this topic produces, and a defended decision not to disclose to a group is part of it.
- What is disclosure letter?
- The actual words sent to affected people, built in a six-part anatomy (what happened, what we got wrong, who is affected, what we have done, what you should do, how to reach a human) and one tone (plain, direct, first-person). It is the second artifact this topic produces and the one a later defense of the file tests for honesty.
- What is available versus disclosed?
- The core distinction of the topic. Information that technically exists but that the affected person will not encounter in the normal course (a hover-over byline, a footer, a terms clause) is available, not disclosed. Only information the person actually learns, where they would see it, counts as disclosure.
- What is quiet non-disclosure?
- Treating available-but-hidden information as if it were disclosure. The pattern behind CNET's "quietly, not in secret" framing, and the failure the topic exists to prevent.
- What is legal floor?
- What a law or contract requires you to disclose (for example, a data-breach notification duty). It is mandatory and it is a minimum, not the full extent of what honesty requires. Verify with counsel whether it actually applies to your incident.
Keep going
This lesson builds Explainability, transparency and contestability, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
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 →