I ran a draft through the AI detector in Substack last week and it told me I was 100% AI. I wrote the piece, ran it through Claude to check grammar, punctuation, and flow and then Pangram scanned my sentences and decided a person didn’t write them.
Of course I had to figure out why. So I fed a few different drafts through trying to find the pattern. What I determined was anything I wrote as pure argument or observation (no story or scene, just me thinking on the page) came back flagged as AI. The pieces that scored as human were the ones where I wrote about myself doing something specific. It seems the detector wasn’t checking whether I was human — It was checking whether I’d left fingerprints.
What bothers me is that none of these tools tell you what they’re grading against. There is no published criteria, No explanation of what “sounds human” is supposed to mean technically. Just a percentage and a verdict. You’re left reverse-engineering your own writing to figure out what tripped the alarm, which is a strange position to be in with a tool that claims to be authoritative about something as basic as whether you exist.
If you’re going to tell someone they didn’t write their own thoughts, you owe them the method.
A lot of competent human writing is generic on purpose. Business emails, professional copy, some substack articles. AI didn’t invent flattened, purpose-built prose, it just got very good at producing it on demand which means the writing humans have always done now reads as AI by association.
I do not like what it implies about where the bar for 'sounding human' has moved. It used to be enough to just be a person writing. Now, apparently, you have to leave evidence. I’m just going to take the compliment and publish anyway — but there's such a stigma around anything AI-adjacent that a flag like this tells the reader not to take you seriously before they've even read the piece. Some people just skip anything marked AI. Watch: soon there'll be a feature to remove that score from your page entirely, for a monthly fee. Like the evolution of how you get a “verified” badge on IG.
People are going to start making their writing worse on purpose to prove they wrote it. If you learn that clean, coherent sentences are more likely to get flagged, you’re going to start writing around the detector. Breaking up sentences that don’t need to be broken, adding awkward phrasing, throwing in a weird sentence structure here and there. Maybe even leaving in mistakes you would normally fix. And that might be the strangest part of all:
A tool designed to detect whether a person wrote something could end up teaching people how to perform being a human.
We’ll add the awkward sentence. Leave the typo. Throw in a personal detail that doesn’t belong. Because some algorithm has decided those are the signs of a human.
At some point, you have to wonder what we’re actually trying to prove.
I wrote this. I’m a person. I shouldn’t have to make the writing worse to prove it.
This piece was written by me and flagged 100% AI by Substack. I’m publishing it anyway.
It was nice to get an extra serving from you with some meat on it, so to speak, after the amuse-bouche you served up earlier in the week.
But I’m a little confused. You allowed AI to complete a piece that you had already deemed finished (and knowing you, I suspect it had already been breathed-over for some time before it was released, even to a machine), and after AI burnished it to the “finish” it favored, then later identified as one of its own, you took exception. What were you expecting? That your “soul” as a writer would somehow shine through, and that the machine would see it and somehow protect it?
You say you “ran it through Claude” as if somehow that act did not count as relinquishing control over the work. If Claude were a human, would you have said, “Claude, I’m going home. Smooth out the grammar and punctuation, would you, and send it off.” Were you not happy with the work, when you asked for a machine’s help? Were you happy (happier?) with the version that Claude produced?
Over-editing has been around for far longer than AI. I’m sure you know from experience that anything creative (and this includes writing, songs, artwork, etc.) can have the “soul edited out of it” by repeated passes, just as the finish can be worn completely through an object by over-polishing.
You say, “People are going to start making their writing worse on purpose to prove they wrote it.” I don’t know if that’s possible; most people can’t write a lick. And for them, AI like Claude may allow their deficiencies to somehow remain less discovered. But it does not mean that they are “writing well”, and for someone of your competency to rely on it is misguided.
You already know that intentionally using “their” instead of “they’re” is not the answer. And, at the same time, I suspect you know already that just not worrying about how a machine identifies it, is.