---
title: "Written by AI"
description: "A label on the message when the words look generated rather than written."
url: "https://openemail.uk/docs/knowledge/safety/ai-written"
area: "Knowledge base"
category: "Safety & authenticity"
status: "live"
---

# Written by AI

A label on the message when the words look generated rather than written.

## Details

- Scored on arrival from the prose itself, with the quoted history and the signature cut off first, so replying to a generated message does not make your reply one. Twenty tells count towards the score and seven against it, weighted and summed: the long dash used six or more times per thousand words, markdown headings no mail client renders, “not just X, but Y”, lists of exactly three built twice over, and sentence lengths that vary less than a person’s do. Typos, lower-case sentence starts, email shorthand, an emoji and a phone signature all count the other way, and can take the total back to nothing.
- The label appears at 60 out of 100 and nowhere below it. No stylistic habit reaches sixty on its own, since it takes two independent tells with no human noise anywhere against them, which is what keeps the mark off people who simply write tidily. There is exactly one exception, and it is not a habit: prose that names itself as a model’s, the “as an AI language model” that gets pasted in unread, is scored at sixty by itself. Between 35 and 60 the verdict still exists, but it stays in the Details panel instead of going on the message, because a hedge printed above someone’s writing reads as an accusation with a disclaimer attached.
- Two kinds of message are declined rather than guessed at, and the verdict says which. Bulk mail is not scored at all. A newsletter is machine-composed rather than model-written, and every tell the two share is evidence of the template, so labelling it would make the label mean “this is an email”. Nor is anything under forty words: sentence-length variance over two sentences is noise, and a long dash in something as short as “yep, Thursday works” is one dash in eleven words. A message that merely arrived inside a template is capped at a maybe rather than refused.
- At 30 and above, on a message that did not arrive inside a template, the assistant reads the words as well, and it can pull the score down by 45 while pushing it up by only 35. There is deliberately no upper cut-off: a high score is the case where we are about to tell somebody a machine wrote their words, and that is exactly where a second opinion able to say “no, a person wrote this” is worth paying for. The skew is deliberate and is the opposite of the one the phishing checks use: a missed label here costs nothing, and a false one tells somebody a colleague’s own writing was machine-made.
- Mail is scored once, at delivery, and the verdict is stored with the message. Anything that arrived before this existed is scored on the way to the screen instead, and that read-path answer is the weaker one and says so: it never calls the assistant, and it sees only the one bulk header a stored message keeps, so a newsletter that declares itself some other way is scored where it would have been declined. The read-path answer is the weaker one and says so: it never calls the assistant, and it sees only the one bulk header a stored message keeps, so a newsletter that declares itself some other way is scored where it would have been declined.
- The switch is in Settings → General, under Mail behaviour, and it is on. Turning it off hides the label and the Details row. It does not stop the scoring. The judgement is made when the message arrives, and a setting read as the screen draws cannot reach back into a handler that ran days ago. Nothing about it is sent to the sender or anywhere else either way.
- No label means nothing stood out, or that we declined to judge. It does not mean a person wrote it.
- It labels, it does not filter. A generated message is not automatically an unwanted one, and mail will not be hidden on the strength of a guess.