Kenneth here.
This week, two of the big AI companies start quietly stamping the text their models write. According to news reports, Anthropic begins watermarking text from more Claude models on Tuesday, and xAI turns on a default watermark for several Grok API models on Saturday. Both point to the EU’s AI Act. You can’t see the mark. Only approved parties can check for it.
On paper, this makes sense. People want to know whether a human wrote the thing in front of them. Teachers want to know. Editors want to know. I get it.
But I have a small problem with the word “human.” Specifically, I’m not sure the watermark thinks I count.
How I actually write
I have ALS. I can’t hold a pen. Typing gets harder every month, and my speech is garbled enough that strangers politely nod and hope for the best.
So I write with a brain-computer interface. I’m participant 21 in Neuralink’s trials, and I dictate by thought into Grok Bot. The ideas are mine. The stories are mine. The bad jokes are, regrettably, mine. The AI helps me get them out of my head and onto the page before I run out of energy, which these days is a real deadline.
I’m a participant, not an employee, and I don’t speak for Neuralink. I’m just telling you how my Tuesday works.
The problem with a stamp
A watermark doesn’t know why you used AI. It only knows that you did.
It can’t tell the difference between a student who asked a chatbot to write a book report and a man who thought every sentence and needed help typing it. To the detector, we look the same. We both get a little invisible sticker that says “not really yours.”
Users are already worried that ordinary edits will get flagged. Now picture the people who don’t have a non-AI option. People with ALS. People after a stroke. People with cerebral palsy, MS, or hands that simply quit. For us, AI isn’t a shortcut. It’s the ramp into the building.
Nobody puts a watermark on a ramp.
What I’m not saying
I’m not against labels. I’m against hype too, and honesty cuts both ways. If a company wants to mark machine-written text, fine. I’d rather know.
I’m not saying these watermarks will ruin anyone’s life on Saturday. I don’t know how they’ll be used yet, and neither does anyone else. That’s sort of the point.
What I’m saying is this: provenance rules are an accessibility question, and right now nobody seems to be asking it. When a school, an employer or a platform decides what to do with a flagged piece of writing, someone needs to be in the room asking, “What if the person wrote this the only way they can?”
What I’d like
Three modest things.
First, say plainly what the mark means. “Passed through an AI tool” is not the same as “written by AI,” and the difference matters.
Second, give people a way to say “this is assistive use” without filling out a medical history.
Third, include disabled writers when these rules get made. We’ve been writing with machines longer than most. We have notes.
That’s the whole ask.
So, is this article AI-generated?
Technically? Probably. Somewhere in here there may be a sticker I can’t see.
But the ideas showed up in my head first. They stayed for coffee. I argued with them, kept the good ones, and sent the rest home.
The machine typed. I wrote.
It counts.
Kenneth