FORM AI-1A (Rev. 2026)

Office of Written Attribution · continuation sheet

If you have been
accused.

Form AI-1 makes the argument. This sheet is the part you take into the room: what to ask, in what order, what each answer is worth, and a letter with the blanks marked. Print it, or keep it open in a tab. Nothing on this page is sent anywhere.

§1

Before anything else

You are not required to prove a negative, and you cannot. There is no evidence that demonstrates writing came from a person — which is exactly why a detector score cannot demonstrate the reverse. What you can do is establish how the accusation was produced and what the people who sell that tool say about it themselves. That is usually enough, because in most cases nobody in the process has actually checked.

If there is a deadline or a hearing date, write it down now and ask in writing for the evidence before it. A request on the record, early, changes how the rest of this goes.

§2

The questions, in order

  1. Which tool was used, and what exactly did it return?

    Not "the detector said it was AI" — the tool's name, the number, and the screen it appeared on. Why: you cannot check a claim nobody will state. A percentage is a different thing from a flag, and a flag on a passage is different again from a flag on the whole piece.

  2. What does that tool's own documentation say about false positives?

    Ask for it in writing. The major vendors state plainly that output is not proof and should not be the sole basis of a decision. Why: this is the strongest single move available to you, and it costs nothing. You are not attacking the tool. You are quoting it.

  3. Was the threshold the vendor's default, or one the institution chose?

    And what does the institution's own policy say about how such a score may be used? Why: a lowered threshold is a decision somebody made, with a consequence they may not have thought through.

  4. Did a person read the work?

    Not the score — the writing. Ask what specifically in it reads as machine-written. Why: an answer in specifics can be discussed. "It was flagged" cannot, and saying so out loud tends to be the moment the process changes.

  5. Has the same tool been run on your earlier work?

    Work from before the accusation, written the same way and never questioned. Why: if it flags that too, the tool is measuring your writing rather than your conduct. This is the single most useful test available and it is rarely run.

  6. Offer your process, and offer it first.

    Drafts, version history, browser or editor timestamps, notes, sources, the messy middle. Why: it is the one kind of evidence that actually points somewhere, and volunteering it reads very differently from producing it under pressure.

  7. Ask for the conversation.

    Offer to talk through the argument, the sources, the choices you made. Why: if you wrote it, you can explain how. It remains the only instrument in this whole business that works.

§3

What the indicators actually measure

Every folk rule behind these accusations sorts writing by care, not by author. Form AI-1 measures twelve of them — em dashes, words like delve, three-item lists, "it is not X, it is Y", signposting phrases, hedging, sentence-length variation, contractions, curly quotes, spelling errors, spacing, sentence length — and then runs the same twelve over Emily Dickinson, Abraham Lincoln, Jane Austen and Herman Melville. They are all referred.

The consequence worth stating in any meeting: careful, formal, edited prose scores highest. That is the writing of someone who revised, someone taught to write formally, and — consistently — someone writing in a second language. A tool that convicts care convicts the people who had to try hardest.

Run your own writing through Form AI-1 and print the result. Being able to show that the same twelve rules refer the Gettysburg Address is more persuasive in a room than any assertion you can make about detectors in general.

§4

Template letter

Replace everything marked. Keep it short and unemotional — this letter is asking questions, not making a case, and questions on the record are harder to leave unanswered.

[Name] [Role, e.g. module convenor / editor / manager] [Date] Dear [Name], I am writing about [title of the work], submitted on [date], which I understand has been flagged as possibly written with AI assistance. I wrote it myself, and I would like to understand how that conclusion was reached. Could you tell me: 1. Which detection tool was used, and what result it returned for this work. 2. What that tool's own documentation states about false positives and about its output being used as sole evidence. 3. Whether the threshold applied was the vendor's default or one set locally. 4. Whether anyone has read the work itself, and what specifically in it was considered machine-written. I am glad to provide my drafts, version history and notes for this piece, and I am happy to discuss the argument and sources in person at any time. I would also ask that the same tool be run on [earlier piece of work, and roughly when it was submitted], which I wrote in the same way and which was not questioned. If it returns a similar result, that would suggest the tool is responding to how I write rather than to how this was written. [Optional, if it applies: English is not my first language. Published research has repeatedly found these tools flag non-native English writing at far higher rates.] I would be grateful for a reply by [date — allow about a week], or before [hearing/deadline] if that is sooner. Yours sincerely, [Your name] [Student/staff number, if relevant]
Print this sheet

Copied from undercoveraiagents.space/kit

§5

If you are the one about to accuse someone

A detector score is not evidence. It is a number produced by measuring formality, and Form AI-1 exists to show that measuring formality convicts Emily Dickinson. Before it becomes a process: run the tool on work by the same person that nobody questioned, and read the writing itself. Then talk to them. If they wrote it, they can tell you how — and if they did not, that conversation will tell you far more than a percentage will.

The cost is asymmetric and worth sitting with. A missed case is a piece of coursework. A false accusation lands on someone who did the work, usually the one who wrote most carefully, and they carry it a long time.

What this sheet is, plainly

It is not legal advice and it is not a detector. It returns no probability and no verdict, and there is nothing here to paste into a disciplinary email. It is a set of questions and a letter, written because the argument on the front page is of no use to somebody who has four days and a meeting.

Free, and free to reuse — print it, hand it out, put it in a student union handbook or a staff guide, no permission needed. Nothing on this page is stored, counted or sent anywhere; there is no analytics on this site.

Form AI-1A (Rev. 2026) · continuation sheet to Form AI-1, a companion piece to undercoveraiagents.site, which does the opposite job: catching the machines that really are pretending. Filed alongside Who decided? and the exchange rate.

Back to Form AI-1