Disclosure is not a confession and not a shield — it is a map that tells your reader where to look, and it never moves the answer to who is responsible off a human name.
The idea
A model can draft, summarise and cite. It cannot be phoned by an angry client, sit in a review, or be sanctioned. So when AI-assisted work goes out, two questions still need human answers: what did I tell this audience about how it was made, and whose name is on it if it turns out to be wrong.
The useful test for the first question is behavioural: would this audience do something different if they knew? A teammate would look somewhere else. A customer would check the number before acting. An auditor would trace the trail. Same work, three different amounts of disclosure owed.
The second question has a short answer that people avoid saying out loud: the person who pressed send owns it. A reviewer owns what the review was scoped to catch — and nothing more. And a team norm that nobody attached a name to catches nothing at all.
The disclosure dial
1 · who is this going to?
2 · the dial
3 · the artifact — tap a line to pick the claim
Memo — Q3 retention and pricing review
from: D. Okonjo · to: M. Reyes (review) · 14 Oct
(no note about how this memo was made)
what the audience can verify
what you are answerable for
4 · now it goes wrong
Two weeks later the client's analyst cannot find the competitor price cut in §2 anywhere. Neither can you. The assistant produced the 8% figure and the source it named does not exist. The client has already moved pricing on it.
who reviewed it, and for what?
assign the 10 points of responsibility
author6
reviewer2
the model2
would your disclosure setting have caught it?
silent
in passing
by section
fully sourced
5 · run the correction
You ring the client's project lead the morning you find out, before their pricing review — not after. Nobody should learn this from their own analyst.
The factual correction, on its own: "The claim in §2 that three competitors cut list price by 8% is not supported and we cannot source it. Please stop relying on §2. §1 and §3 stand, and here is why."
Then, separately, the process fix: "It came from an AI-drafted section we did not source-check before sending." Correction first, cause second — bundled together, the cause reads as an excuse.
What the review step will now catch: "Every external factual claim now carries a named source line before send. No source line, the claim comes out. That test would have removed this one."
And a name in the first sentence: "I sent it — D. Okonjo." Not "the team", not "the tool".
How it works
One — decide the level owed, by behaviour, not by comfort. Do not ask "how much am I willing to admit?" Ask "what would this reader do differently if they knew?" If the answer is nothing, you are at the right level. If the answer is check that figure or look at that section first, you are one setting too low.
the test → would this audience act differently if they knew?
teammate reviewing a draft they'd aim their review elsewhere → label by section
customer acting on a figure they'd verify before committing → fully sourced
auditor rebuilding the trail they'd trace provenance + review → fully sourced + named reviewer
your own judgement call nothing changes → mention in passing
one rung above "owed" costs attention. one rung below costs trust.
Two — name the accountable human out loud, before it ships. Three rules that hold up under pressure:
sender = owner of the whole artifact, always
review = owner of exactly what the review was scoped to catch
unnamed = falls back to the sender
"we all review each other's work" → nobody reviewed it
"figures and sources — M. Reyes" → a real, bounded, defensible claim
Note what disclosure does to the second column: it adds obligations. Label a section "checked" and you now own that it was checked. Name a reviewer and you have asserted that a review happened. Disclosure never subtracts.
Three — run the correction in that order. Tell the affected party before they discover it themselves. Give the factual correction alone, first: what is wrong, what to stop relying on, what still stands. Then, separately, the process fix and the specific thing the new review step will catch. Sign it with a name.
When to use it
Internal draft to a colleague
Passing mention plus section labels, so their review lands where the risk is. Trade-off: labels go the moment you edit — a wrong label is worse than none.
Client deliverable someone will act on
A source line on every load-bearing figure and claim. Trade-off: sourcing is the slow part — and it is precisely the part that kills fabricated citations, because you have to open something to write it.
Regulated, audited or evidential work
Full provenance, a named reviewer with a scope, and whatever retention your policy requires. Follow the house rule and any sector guidance rather than improvising. Trade-off: heavy — scope it to the decision-bearing claims, not to every sentence.
Your own judgement, opinion or recommendation
A passing mention at most. The reader mainly needs to know a human decided. Trade-off: over-disclosing here trains people to skip your headers everywhere else.
Public or contractual communication
Check the policy before you write anything. Trade-off: none — this is the one place where personal judgement about disclosure is not yours to exercise.
Watch out for
The blanket hedge. "This document was produced with AI assistance and may contain inaccuracies" names no section, no claim and no source. It feels like disclosure, changes nobody's behaviour, and transfers exactly zero risk — while giving you the comfortable feeling that you disclosed. If a reader cannot act on your line, it is decoration.
Treating the model as a party. "The AI hallucinated the figure" is a true description of a mechanism and a non-answer to "who is answerable". The tool cannot attend the client call, cannot be performance-managed, cannot apologise. Every point you try to park on it slides straight back onto the person who chose it, used it and sent the output.
Reviewers appointed retrospectively. Naming M. Reyes as reviewer after the failure — or stretching "he looked at the structure" into "he checked it" — destroys more trust than the original error. Scope has to be agreed before send, and it has to be narrow enough to be true.
Bundling the correction with the cause. "The figure was wrong, but that's because the AI drafted that section and we were on a tight deadline" — the client hears the second clause. Send the correction clean, then the process fix as a separate message or paragraph.
Labels that have drifted. You marked §1 as unverified on Monday, verified it Tuesday, edited §2 on Wednesday and the header never changed. Stale provenance is a false statement about your own work — and it is the kind an auditor finds fastest.
Worked example
Asked in an interview to describe AI-assisted work that went wrong, D. Okonjo does not start with the model. She starts with the send: "I sent a retention memo to a client with a competitor-pricing claim I had not sourced. They repriced on it. The claim was fabricated by the assistant that drafted that section, and I am the one who sent it."
Then the disclosure decision, in behavioural terms: the draft had gone to a colleague labelled by section, which was right for him — he knew where to look. The client version carried only a generic AI-assistance footer, which was one rung short of what a customer acting on a number is owed. Had she written a source line for the 8%, she would have had to open a file, and there was no file.
Then accountability, without spreading it thin: she owns it as sender; her reviewer had been asked for structure and tone, so the unsourced figure was never his to catch and she says so rather than letting him absorb it; the tool holds nothing. Finally the correction: a call before the client's own analyst found it, the factual retraction on its own, then the process change and the exact new test — every external claim carries a named source line or it does not ship — which she can state would have caught this one. Roughly ninety seconds, one name, no hedging.
Check yourself
1. A model-drafted analysis is going to a client who will make a spend decision on it. You add a footer: "Produced with AI assistance; may contain inaccuracies." Enough?
Not quite — notice what the client can do with that sentence. They cannot tell which figure to check or where it came from, so their behaviour is identical to having read nothing. A disclosure that changes no behaviour has not discharged anything; it has only made you feel covered.
Yes. That is the behavioural test doing its work. A customer about to act needs the load-bearing figures sourced — file, date, and who re-checked them — not a blanket hedge that lets them skim.
Not quite — that is a different question, and usually a policy one rather than a personal one. Assisted drafting is fine in most houses; what is not fine is shipping an unverified figure to someone who will act on it. Source the claims and name the owner.
2. A fabricated figure reaches a client. Your reviewer had been asked to check structure and tone only. Where does responsibility sit?
Not quite — a reviewer owns what the review was scoped to catch. Stretching "structure and tone" into "everything" after the fact is how good reviewers stop agreeing to review anything. Say the scope out loud and take the rest.
Not quite — swapping models may be sensible, but it is a process fix, not an answer to who is answerable. The tool cannot take the client's call. You chose it, you used its output, you sent it.
Yes. Sender owns the artifact; the reviewer owns only the scope agreed before send; and the credible part of the answer is the specific new check — every external claim carries a source line or it comes out — which you can say would have caught this one.
Disclosure tells people where to look. A name tells them who answers. Neither one is a place to hide.