At first I thought this meant it would insert invisible characters, and I figured a fast typist could get around this by just re-typing Claude’s output. Not so fast, sucker! said Claude.
What Claude actually does is uses patterns of words that it can automatically recognize. It’s more sophisticated than words like “quietly,” punctuation marks like an em-dash, or annoying phrases like “This hits different.”
Instead, Claude creates patterns in its choice of words that will cause the block of text to be recognized as AI. Supposedly it won’t affect the text’s meaning or readability.
Here’s the thing. (That was me joking, not an AI writing this…) The feature has been rolled out already to comply with the E.U.’s deadline. But you can’t actually use it yet to detect A.I.-generated text. Claude will release tools to do that later. Also, presumably, pre-2026 slop will survive as is.
According to (Tim Parkin)[https://www.linkedin.com/in/tparkin/], this means that if you write a brilliant piece and ask Claude to do a light copy-edit, your work may be labeled as AI. People writing about it caution that this doesn’t mean Claude will own the copyright to your work, but I think the main issue is perception: as soon as one of your pieces gets labeled slop, the reader won’t trust it as much.
Also, presumably, if I put out a bunch of A.I.-generated stuff last year, the slop-o-meters won’t pick it up because the A.I.’s weren’t adding watermarks then. Fun stuff!
People are already saying they’ll fool the system by translating their text into another language and then translating the translation back into English. Isn’t that how every user manual is written?
In the meantime, I’m going to keep minimizing my use of em-dashes and I’ll try not to do anything quietly.
