I use AI to write everything I publish here. Substack now grades that.
Interfaces create norms through what they measure, what they display, and what they make us explain.
If you found your way here because of an article using me as an example of someone “hiding” their AI usage, hi! This publication is pretty much only about human–AI collaboration. Imagine how fun it would be if I argued for using AI to augment human creativity and still went, “But don’t you dare use AI as a tool for writing your Substack articles.”
When I was around nine years old, I took my little pink notebook to the games of our local soccer team and wrote down what I saw: who scored, what the crowd was like, any injuries or substitutions, and my opinions on who should take over as head coach the following season.
I don’t remember who I imagined would read it. I’m not sure I imagined a reader at all. Something was happening, and I had a notebook, so obviously I was going to turn it into a story.
As a kid, I was always making something. I wrote books and made magazines, ran little businesses and experiments, picked up photography, and taught myself to build websites. I kept making things as I got older. I even wrote a few articles for our printed university paper. But eventually everything became attached to something: a course, a job, a brief, an outcome. I could still create when someone needed something from me. I had forgotten how to make things just to see what they might become.
And yes, this essay is about writing with AI.
Because writing with AI helped me find the playfulness I had as a kid. The conversation keeps me interested enough to finish, and the pushback helps me see what I actually want to say. As a non-native English writer with a neurodivergent brain, AI also helps me reach thoughts I can feel before I can explain them.
The notebook gave my thoughts somewhere to go. AI does too, except now the notebook talks back.
Last night, I read that Substack has turned on AI detection. It runs on Pangram and gives readers a percentage of how much of a text is human-written or AI-assisted. Writers can also add a “How I make this” statement that people will see when they scan our work.
The announcement is called Against Claudefishing.
I was annoyed before I had finished reading.
This was one of those moments when the only way out of the annoyance was to write my way through it. So I opened AI and started talking. Which is, inconveniently for the detector, exactly how this essay got made.
How I actually write with AI
I’m a verbal processor, and I need a lot of reps to find the right words. My thoughts arrive as a feeling that something is off: an irritation, two things that seem connected though I can’t explain why, or a sentence that makes my whole brain fixate on understanding it.
So I write messy. I say the thing badly, hear it back, and realize that isn’t what I mean. I correct myself, contradict myself, and follow an argument until I get bored. Then I take it one step deeper, and sometimes another one after that.
AI can keep going for as long as I need. It holds the thread while I orbit what I am really after and gives me a version of the thought to push against. Sometimes it finds the word I couldn’t reach. Other times it gives me five paragraphs that are technically correct and make me feel absolutely nothing, which is useful too. Now I know what I don’t mean.
There is never a finished essay or argument sitting inside me, waiting to be typed out. That’s just not how my brain works. Talking it through is how I understand the experience. I often don’t know what I think until I’ve said it several different ways and felt which version is true.
I don’t make these essays the way I used to. My first ones were built almost entirely inside the conversation. I arrived with a feeling, talked until something took a shape I recognized, and the text came out of that. Now I often arrive with the shape already. I know what I want to say, and the conversation exists to break it.
That doesn’t mean I use AI less nowadays. I write more to it now than I did then. The volume went up and the job changed.
A detector might call that progress. But it would be measuring what changed in the final text, not what got better in my thinking.
Somewhere in all those conversations, I started making things for no reason again. A note because one sentence amused me, an essay because an argument wouldn’t leave me alone, a duck where no duck was needed. At 34, with two kids, I’ve found my way back to play.
That is the part of the conversation about AI I rarely recognize. People talk about efficiency, cheating, replacement, and whether the sentences count as yours. I’m over here feeling like someone quietly unlocked the room where I used to make things.
What the percentage gets wrong
I understand the problem Substack is trying to solve. The internet is filling with mass-produced text about nothing, and readers want to know whether there is a person on the other end.
But the detector arrives at the end of everything I just described and uses AI to look for traces of AI, which is almost too funny.
It doesn’t see the feeling I couldn’t name, the private conversation that went in circles, the arguments I rejected, or the sentence I rewrote because it sounded good but didn’t feel right. It sees the finished language and looks for a machine.
The percentage will tell you something about the surface of the text. It won’t tell you how true it is.
I’m not arguing that Pangram will get the percentage wrong. Even a perfect score would answer the wrong question. Readers want to know whether a person thought about what they published, cared about it, and meant it. How much AI touched the language cannot tell them that.
Substack knows this. The announcement says the detector cannot tell whether real care went into the work. But that caveat lives in the announcement, while the percentage lives beside the writing. Guess which one people will pay attention to?
Once that percentage sits beside the writing, it stops being neutral information. The writing has been handed a grade.
A low score is the clean result. A high score needs context. Turning detection off becomes its own kind of statement. Before anyone reads my “How I make this,” the design has already suggested a correct way to make writing: by hand, with as little machine in the sentence as possible.
Maybe Substack does not intend to set that rule. But interfaces create norms through what they measure, display, and make us explain.
AI is part of both my thinking and my writing. Using it feels like an expansion of what I can do, not a replacement for the part I bring. I’m still the one deciding what feels true, what gets rejected, and what I finally publish.
I use AI openly. I also know there is a difference between text produced because someone needed more content and writing that began with something a person couldn’t leave alone. I can feel that difference as a reader. I don’t think a percentage can feel it for me.
So here is my “How I make this” statement:
🦆 I write by talking to AI. I begin with a feeling, a question, or something I can’t quite explain, then go over it again and again. I read what comes back, notice what feels wrong, push against it, and keep shaping the writing until I recognize myself in it. That conversation is where I find out what I think. AI is a vital part of how I make this.
That is the short version.
The longer version is this whole publication: every thought I followed because I wanted to know where it went, and every draft that helped me understand the experience too.
I still don’t know exactly why I stopped making things for the pure joy of it.
I know that last night I got annoyed by a product announcement, opened AI, talked until I found the thing underneath the annoyance, and made an essay where there wasn’t one.
I can’t wait to see what score it gets.
The detector can measure whether AI was involved. It cannot measure whether the thinking was. We keep building the metric that is easy instead of the one that is true.
As a fellow non native speaker and neurodivergent person, I agree with everything you said. I use AI for writing the same way you do.
By the way, here's a funny tidbit: when I tested Pangram (as web publisher), it flagged my own non-AI writing as AI. When I used zero AI. I'm pretty sure that yes, I sound a bit like Claude now because I spend so many hours talking to it. It rubs on you.
But yes, the whole AI vs. Human content is stupid and outdated. Either the content is good and interesting and engaging. Or it's not. How it came to be doesn't need to be flagged.