Flagged as AI on Substack

Substack announced it in Against Claudefishing. Readers can scan notes, replies, comments and posts and see Pangram’s estimate of how much of the text was written by hand and how much with AI assistance. It works on text longer than 100 words, published from that day on, and shows its analysis only to whoever asks for it.

The same announcement says the scan “is not perfect”, and that Pangram can tell whether AI was used to make a text, not how much human care went into it. A scan is an estimate from the finished words. It does not know how they were written.

Substack gives writers three things, by its own account:

• –Report and remove a scan of your own work that you believe is mistaken.
• –Add a “How I make this” statement that explains your process to readers.
• –Run Pangram on your drafts before you publish.

The last is worth using with care. Rewriting a sentence until a detector stops objecting to it bends your style towards whatever the detector happens to like, and writers have already said they feel pushed to give up the em dash for this reason. Your voice is what your readers subscribed to.

A detector looks for text that resembles what a model produces, and clear, even, carefully edited prose resembles it more than a rough draft does. Writing in a second language, a formal register or a tight format can score the same way. Why detectors flag human writing goes into it.

A line at the end saying you wrote it is a claim, and anybody can type it, including somebody who did not. What readers can weigh is a record of how the post was written.

iwrote keeps one. You draft the post on iwrote, where the page records how the words arrive: what was typed, pasted in, deleted, and where you paused. When it is done you certify it and copy a line with a short link to put under the post on Substack. Whoever opens the link sees how long it took and how it was written, and can check the record without taking iwrote’s word, or yours.

It covers writing done on iwrote from the first word, so it cannot vouch for a post already published. For those, what you have is your drafts and their history; see how to prove you wrote it.

See what a reader gets, or start your next post on iwrote.

If you answer, keep it short: you wrote it, here is how, and here is the record if there is one. An argument about a score is rarely won, because neither side can see inside the detector, and a reader who sees a calm answer with something to check tends to remember that more than the score.