why detection fails · what you can actually claim · the four-line account
Two questions get collapsed into one, and they have different answers. Whether a tool was used cannot be established by anybody, including you six months later. Whether the work holds can be established, in four lines, by the person who did it.
How it was made, and whether it holds. Almost every argument about AI and honesty is people answering the first question at each other while the second one goes begging.
An account of what you checked survives being questioned, and a claim about how the words arrived does not survive the first person who asks how you would know.
People look for: AI disclosure, AI detection limits, evidence records, audit trails, citation checks, and accountable approval.
Everything in the top half is a claim you cannot support if pressed. Everything in the bottom half you can, in a sentence, with a document behind it.
| The claim | Can it be established? | By what |
|---|---|---|
| No AI was used in this document | No | Nothing. Nobody can show it, including you, once time has passed |
| This was written by a person | No | Detection clears real machine text and accuses real people. It settles nothing |
| AI was used only for editing | No | A boundary nobody can locate afterwards, including the author |
| Every figure was checked against the source named beside it | Yes | Open the source. Two minutes |
| Every name was confirmed with the person | Yes | The messages exist |
| This claim is not verified, and here is why | Yes | It is a statement about your own work, and it costs you to make it |
| These three decisions were mine, and here is the reasoning | Yes | The record from day 20, written at the time rather than afterwards |
What the piece claims. What you checked, and against what. What you could not close, and what would close it. Who decided the things that were judgement rather than fact. Four lines, written while the work is fresh, and they answer every reasonable question anybody will ask — while the question everybody argues about goes politely unanswered, because nobody can answer it.
Defensibility comes from evidence and named verification, not from guessing how prose was produced or attaching a tool name without context.
Somebody has written a rule, or nobody has. Go and find out which, and read the actual words rather than the summary somebody gave you in a meeting. Calibrating to what colleagues seem to be doing is how people end up on the wrong side of a policy that was sitting in a shared folder.
It fails in both directions, and the second failure is the serious one: careful, plain, well-structured prose is exactly what gets flagged, so the writers most likely to be accused are the ones who took the most care. Never rest a claim on a detector, and never accept one resting on you without asking what its error rate is.
Reconstructed six months later, an account is a reconstruction and everybody can tell. Written the afternoon you finished, it costs four lines and it is the difference between somebody who checked and somebody who says they checked.
The team answered with four concrete lines: assistance used, sources supplied, checks performed, and final approver. The account was useful because each statement could be examined.
A detector score was treated as proof until reviewers asked about false positives and direct evidence. The decision was restarted using process records and the student's actual work.
A published number was challenged after its source changed. The saved source version, calculation, review note, and approver allowed the organisation to explain and update it without guessing.
Defensible work has traceable sources, named checks, recorded limitations, clear decision ownership, and an account of where assistance was used when that fact matters. The aim is not to prove a document's hidden production history; it is to support the claims and decisions people rely on.
No detector can establish authorship reliably from prose alone. False positives can accuse careful human writers, while edited machine output can evade detection. Treat a detector score as an uncertain signal, never as proof, and use process records or direct evidence for consequential decisions.
State the material activity rather than naming a tool without context: what assistance was used for, which sources or inputs governed it, what a person checked, what remained uncertain, and who approved the result. The detail should match the audience, risk, contract, and applicable policy.
Citations make individual claims inspectable when they point to relevant, accessible sources and the author has checked that those sources support the claim. A long reference list does not prove the analysis, and an assistant's citation must be opened and verified before it carries weight.
Keep the source materials or versions, key instructions, material outputs, model or tool version when relevant, verification performed, unresolved limitations, changes made by people, and final approver. Retention should follow the work's risk and privacy rules rather than storing every conversation indefinitely.
You cannot prove how it was made. You can prove what you checked, and only one of those was ever the question.