This week, I typed out a post entirely in my own words and ran it through an AI checker. The checker scored it as 100% AI-generated, and that happened more than once. AI content detectors are broken, and this is going to hurt the wrong people.
What happened when I tried to publish my own words
I write across a handful of platforms for myself, my own business, and for clients. This week, I used Substack’s AI detector for the first time in a serious way. I wanted to test it out. I typed a post manually, word by word, typing every sentence directly from my brain; I didn’t use copy-paste or AI drafting. Substack’s checker scored it at 100% AI-generated. The score stayed flat at 100% throughout.
I ran my text through a couple of tools to see if the result would change: a “ghost checker” and a slop-trim tool built to clean up AI-sounding phrasing. Both of them changed my handwritten text, and it still scored as 100% AI-generated.
AI-checker tools measure sentence rhythm and evenness. They treat smooth, consistent sentences as a sign of AI generation. Writing in clean, complete sentences produces a high AI score, regardless of who wrote it.
AI content detectors are broken, and clear writing gets the blame
The false positive is one problem. What concerns me more is what it implies about who gets believed. I use AI tools like Claude and ChatGPT to take my own ideas and put them into words other people can follow. The thinking is mine. The tool provides the structure.
When a detector labels a handwritten paragraph as machine output, the label carries a message: clear, direct writing looks suspicious to the system. Training an entire industry to treat clarity as suspicious is a real problem.
Why neurodivergent writers are affected most
I’m neurospicy. Getting an idea out of my head and into words I can send to someone else takes real effort. AI helps me articulate ideas I already have. Many neurodivergent writers work the same way. The thinking is theirs. The tool translates it into a form a reader can follow on the first try.
When detectors treat precise, well-organized writing as a sign of AI use, the writers affected first are the ones already told for years that their natural communication style needs correction. A percentage score now attaches to that judgment and gets treated as evidence.
The double standard between Substack and LinkedIn
Substack lets you turn the AI checker off. LinkedIn keeps its checker running at all times, as far as I can tell. Publishing on Substack means you control whether that score exists. Publishing on LinkedIn means a permanent, invisible score attaches to your content, fixed and outside your control.
That’s a major inconsistency: two different sets of rules for the same behavior, decided by which platform owns the checker.
I spent the same week teaching people to use AI properly
The same week that detectors scored my own words as fake, I ran a live session teaching a group how to build a real tool with AI. We built a tiny web app together called The Translator: one text box in, three outputs out, plain language, professional version, and the reverse-engineered prompt behind the original text. We kept it small on purpose, so people could finish it in one sitting.
That session also covered a point I think is underrated: being neurodivergent and literal is a genuine advantage when prompting AI. You say exactly what you mean, which is exactly what these tools need to work well. The same trait that gets flagged by a content detector is the trait that makes you good at directing one. That connection deserves more attention.
We also talked about keeping AI skills and prompts in a portable place, like GitHub, rather than inside one company’s app. That way, the same instructions transfer directly when you switch tools. These detectors get things wrong with confidence, which makes portability and control over your own workflow important.
Part of that conversation was about switching between AI tools less often. Several people in the group described moving from one AI tool to another, then feeling pressure to move again every time a new one launches. The collective conclusion was to keep one primary tool and try new ones on the side, rather than rebuilding a workflow from scratch every few months. Given that a checker can score identical writing as 100% AI generated one week and pass it the next, a stable workflow that you understand and control matters more than chasing whichever tool is popular that week.
One more decision I made this week
Separately, I made the decision to pause a side project I had been running with a collaborator. We had four hours of meetings scheduled every week, and the project stayed exactly where it started. Recognizing that a project has stalled and stopping it counts as a real result, the same as finishing one. We’re keeping the accounts active in case the energy returns. It was a straightforward decision.
What I’m doing about it
I’ll keep using AI to help me write, and I’ll keep disclosing it when that matters. I’ll treat a detector’s percentage score as a rhythm measurement rather than a truth measurement. These tools match patterns in sentence rhythm. Right now, they produce confident, wrong scores that penalize exactly the writers who deserve more consideration: people who use AI to say what they already think, more clearly than their own first draft allows.
When a detector scores your own words as machine-written, the score reveals a limitation in the detector. Your writing remains yours.