What to do if you are accused of using AI to write

The instinct is to explain that em dashes are normal, that you have used them for years, that the character predates ChatGPT. All of that is true, and none of it helps, because the accusation was never really about the dash. It was about a hypothesis, and the dash was the visible part of it.

Gather whatever exists, in this order of usefulness:

1. Version history. Google Docs and Word both record edit timestamps. A document built over nine sessions across three weeks is very hard to explain as a single paste.
2. Earlier drafts. Even a rough one. Drafts show direction of travel: what you cut, what you moved, what you got wrong first.
3. Notes and outlines. Handwritten pages, a notes app, a messy bullet list. Anything that predates the final text.
4. Sources with annotations. Highlights, margin notes, a reading list. This doubles as evidence of how you worked.
5. Messages where you discussed it. A Slack thread, an email to a supervisor, a question you asked a colleague.

If none of this exists, say so plainly and shift to the substance instead. A clean, honest "I wrote it in one sitting and I do not keep drafts" is more credible than a manufactured trail.

This is the strongest move available, because it exploits the actual difference between generated text and written text: the author can explain the choices. Invite the questions and mean it.

I understand the concern, and I would rather resolve it directly than argue about software. I can walk through any section you pick: why I structured it this way, what I left out, and how each source led to the claim it supports. I can also show my drafts and the notes I worked from. Would that help?

• Expect questions about the things only you would know: the part you cut, the contradictory source you dismissed, the reason you chose one example over another.
• Do not prepare a script. A rehearsed answer to "why did you leave that out" sounds exactly like a generated answer.
• Keep it warm. The person raising this usually has a policy they are required to follow, not a personal vendetta.

Most institutions treat a detector score as a signal to investigate, not as a finding. That is in your favour, and it also means the process is about the record more than the number.

• Ask, in writing, what specific evidence is being relied on. Policies usually distinguish between a detector output and an actual finding, and asking forces the distinction.
• Ask which policy clause applies, and request it in writing. Vague accusations are hard to answer.
• Point out, politely, if you are writing in a second language. Detectors are documented as flagging non-native writers at much higher rates, and many institutions now acknowledge this.
• Bring the record from Step 2 and offer the conversation from Step 3 at the first meeting, not as a last resort.
• Keep every message. If it escalates, an appeal is built from the paper trail you keep now.

Commercial accusations are usually about trust rather than rules, so the response is about process as well: show the drafts, offer the walkthrough, and demonstrate that you can reproduce the work. Then fix the underlying situation, because a client who thinks your work looks machine-generated is a client who is about to leave.

The self-check is a reasonable instrument here, and unlike a detector it reports something actionable. So is running the finished text through the AI dash remover before delivery, which takes a few seconds and removes the most visible trigger.