What AI actually changes about writing
I write these field notes myself. I mention that because it has become a slightly strange thing to say. A machine can now produce a clean, confident paragraph on any subject in about a second, which makes it tempting to conclude that writing is a solved problem and the rest of us can put our pens down. I’ve watched a lot of players reach that conclusion. Their output gets faster and, quietly, forgettable.
In the daylight layer I’m a cofounder of Wistkey, so I’m the last person to tell you not to use the tools. Use them. But it’s worth being clear-eyed about what they change and what they can’t, because the two get muddled.
Writing was never only the words
The sentence on the page is the output of writing. The actual work is the thinking that produces it — noticing you don’t quite believe your own first draft, finding the example that makes the idea land, cutting the clever line that doesn’t survive scrutiny. That struggle isn’t a tax on writing. It is the writing. It’s where a vague feeling turns into something you can actually stand behind.
AI can write the sentence. It can’t do the thinking the sentence was supposed to force.
When you hand the whole thing to a model, you get fluent text and skip the struggle — which means you also skip the part where you find out what you actually think. The prose reads fine. It just doesn’t come from anywhere, and readers can feel the difference even when they can’t name it.
What AI genuinely helps with
None of this is an argument for doing it the hard way on principle. AI is genuinely good at the parts that are friction without insight:
- Beating the blank page — a rough draft to react to is easier than nothing to react to.
- Editing — tightening, catching repetition, flagging where you’ve lost the thread.
- Reframing — “say this three other ways” when you’re too close to your own phrasing.
- The mechanical — formatting, summarizing, adjusting length or tone for a specific reader.
What it can't hand you
It can’t give you something to say. It can’t tell you which of ten fluent options is the true one — that’s taste, and taste is built from having done the thinking before. And it can’t give you a voice; a voice is just the residue of a particular person deciding, over and over, what to keep and what to cut.
So here’s the rule I use, in the daylight layer and in this odd little room: let the machine remove the friction that’s only friction, and keep the friction that’s thinking. Draft with it, argue with it, edit with it. Just don’t let it do the one part that was the point — deciding what you actually mean. That part still has to be yours, or the words belong to no one.
