Last month I published a LinkedIn post about a side project. It read fine to me. A browser
detector put it around 60% AI. Gemini, asked to judge it the way Grammarly would, said 85-90%.
I deleted it and rewrote the thing from scratch, keeping the same facts. The rewrite scored
0-10% on Gemini and I felt pretty good for about an hour, until I fed the identical text to
two other models. Grok landed somewhere around 65-75, and ChatGPT thought 70-85 was fair.
Same words, a spread of roughly 70 points depending on who you ask.
After that I mostly stopped caring what the detectors said, because human readers kept catching
my drafts anyway, and they were more consistent than the tools: when I collected what people
actually pointed at, it was the same short list of patterns every time. That list became a
checklist I now run before publishing anything. It started at 7 rules. By the time I sat down
to write this post it was at 21 (more on that number at the end).
Some rules you already know: "it's not X, it's Y" contrasts, em-dash overload, "In today's
world" openers. The ones that surprised me:
Gestured specificity. Phrases like "a real number" or "an actual example" with no number
and no example anywhere in the text. Reads as concrete on a skim, and is exactly as empty as
"many companies struggle with this". My deleted post had two of these.
The over-resolved arc took me longest to see because no single sentence trips it. Label
each paragraph with the one job it does: setup, numbers, cause, fix. If every paragraph does
exactly one job, nothing overlaps and nothing dangles, that cleanliness is itself the tell.
Off-the-cuff writing leaves threads hanging somewhere.
Then there's the way numbers get reported. "A said 8%. B said 71%. C said 82%." Three data
points in identical syntax read as manufactured evidence even when every number is real,
which is why the detector scores at the top of this post don't line up in a neat row.
And the closing question. Every draft I wrote wanted to end with "What's your experience?"
Readers learned to see that as template long before I did. A post can also just stop.
The worst thing the checklist ever caught wasn't stylistic though. I found two published
comments from my own account, under other people's posts, that said some version of "this
happened to me too." It hadn't. The drafts passed every style rule I had at the time, sounded
perfectly human, and claimed a personal experience that didn't exist. That's not a tell,
that's a lie, and a reader who spots it is right to write off everything else you say. It
became its own rule, the only one where the fix is "ask the author whether this actually
happened", and it's the rule I'd keep if I had to drop the rest.
I packaged the checklist as a Claude skill so I stop having to remember it: drop it into
~/.claude/skills/, feed it a finished draft, get back a table of line → pattern → fix. Free
on GitHub, English and Russian: https://github.com/aragossa/ai-tell-detector. It's the one
skill I give away from a paid pack of nine, because it's the one everybody seems to need. Fair
warning from the skill's own hard rules: passing the checklist doesn't mean a detector will
score you 0%, since detectors can't reliably tell "human draft, AI-assisted edit" from "fully
generated" anyway. The goal is that a reader can't name the template you wrote from.
About that "21". A draft of this post scored 44% on Grammarly, so I showed it to ChatGPT and
Gemini and asked where a classifier would push back. They independently flagged the run of
bolded term-definition blocks in the middle ("looks like a generated taxonomy"), a punchy
"So X is noise. What isn't noise:" pivot I was honestly proud of, and the tidy caveat
paragraph I'd slotted right before the ending. All three went into the checklist, which is
now at 24, and I rewrote this post against the new rules before publishing. I doubt it's
done. Readers, and apparently other models, keep finding new ways to catch me.
Top comments (0)