I dictated the first version of this article. When I clicked a button to turn my speech into text, the tool told me it had also made some edits. I opened the transcript and found a cleaner, more coherent version of what I had said.
The feature was useful, though it also skipped a step I hadn’t meant to skip.
I started with a question about writing versus dictation because writing is slow. You have to find the keys, fix typos, move sentences around, and deal with the small failures of hand-eye coordination that come with typing. Dictation clears away much of that trouble because you talk and words appear.
But the more I worked through the question, the less interesting that comparison became. Dictation is just one way to capture thought, so the larger question is what work AI should remove from writing and what work the writer still needs to do.
My answer is that we should remove as much mechanical friction as we can. At the same time, we should protect the conceptual friction that helps us think. The point of writing should be to work through an issue. Constructing sentences and paragraphs is a means to that end, and AI can now carry much more of that burden.
The trick is keeping it from carrying the thinking too.
Two kinds of friction
There are two kinds of friction in writing: mechanical and conceptual.
Mechanical friction is easy to spot. It’s the effort of typing, correcting spelling, cleaning up grammar, and moving paragraphs into place. It slows down the transition from thought to text. Sometimes it slows it down a lot.
I see little value in most of this effort because missing a key doesn’t make an idea wiser. Fixing a typo doesn’t deepen an argument. If dictation can capture an idea faster, I want to dictate, and I’m happy to let AI repair a sentence without changing its meaning.
Conceptual friction is different. It appears when I reread an idea and realize it doesn’t fit. I may find that two claims clash, or that an example points somewhere I didn’t expect. Then I have to decide what I mean. That work can be annoying, but it’s also where much of the value lies.
The two forms of friction used to arrive as a package. A writer had to clean up the prose, which forced the writer to spend more time with the ideas. Mechanical work created an occasion for conceptual work, though the mechanics had little value on their own.
AI lets us split the package apart. That’s a gain, provided we know which half to keep.
Research by Veerle Baaijen and David Galbraith helps explain the distinction. Their study links discovery through writing to two separate processes. One is spontaneous sentence production, which can generate new content as a person writes. The other is revision of the larger structure, which can change how the writer understands the ideas as a whole.
Dictation seems well suited to the first process because speaking freely brings related ideas to mind, often faster than typing can record them. The second process depends on what happens later. If I review the transcript and reorganize its ideas, I still have a chance to rethink them, but I lose that chance when AI does all of the work.
This is why the writing-versus-dictation frame led me astray. Imagine a stenographer who could record every word as fast as I could speak without doing any of the editing. The speed of capture wouldn’t prevent later discovery because I could still read the transcript, question it, and rebuild its structure. The real risk enters when AI handles the editing as well.
The clean-up trap
A common AI writing process is simple. A person writes or dictates a rough stream of thought, then asks AI to organize it and turn it into a finished piece.
The result may be easier to read and may even express the original ideas well. But the writer can move from rough thoughts to polished prose without ever wrestling with the argument.
I’ve noticed this in my own use of dictation. A long transcript appears, and I feel as if I already know what it says. After all, I just said it. I am less eager to read the whole thing again, especially when AI is ready to do the cleanup.
At first, I blamed dictation for that distance, but the same thing can happen after typing or writing by hand. AI automation during editing creates the distance because it removes the need to return to the ideas.
Dictation may increase the temptation because spoken thoughts tend to be wordy and loosely ordered. A transcript can look like a small disaster, even when it contains good ideas, so handing the mess to AI feels like the obvious next step.
That’s where a useful tool can quietly take over the most important part of writing. AI may fix the sentences while also deciding what belongs together, which points matter, and what should come first. Those conceptual choices could have led the writer to notice something new.
The key difference is between asking questions and making proposals. A question sends me back into my own thinking. A proposed outline begins to settle the structure for me. Even a neat summary can smuggle in an order that I didn’t choose.
So the answer cannot be a ban on AI editing. That would keep plenty of useless friction and give up a powerful tool. We need a better order of operations.
A writing process built for discovery
I’m testing a process that gives AI different jobs at different stages. Early on, it should help me encounter my ideas. Later, it can help turn those ideas into clear prose.
First, I dictate or write freely without starting from a full outline. I try to capture the issue, the claims that seem important, the stray examples, and the thoughts that may or may not belong. Speed matters here because every delay is another chance to lose the next idea.
Then I use AI as a mirror and a questioner. I want it to help me see what’s already present without deciding what the piece means. It can ask which ideas I have mentioned, where two claims seem to conflict, what remains unresolved, and which questions the material raises. The goal at this stage is to send the work back to me.
I respond by adding more raw material. Some questions lead nowhere, while others reveal a missing distinction or a claim I had not known I was making. I append those thoughts and repeat the review without polishing the growing transcript yet.
At some point, the task shifts from finding ideas to organizing them, and I think the writer needs to own this part. AI can keep asking questions, but the writer should decide what belongs together, which idea leads, and what final conceptual shape the piece should take.
Only then do I ask AI for a full edit. It can fix grammar, cut repetition, build clean transitions, and turn fragments into paragraphs. I review that draft and ask for changes. New ideas may still appear, and I can work them into the piece before another pass.
How do I know when the discovery stage is finished? I probably don’t. There’s no meter that reaches 100 percent then chimes. For now, the test is whether I feel like the central questions have been answered and whether a reader finds the result satisfying. That’s subjective, but so is much of writing.
I also can’t prove that this process produces as much discovery as writing and revising every sentence by hand. Baaijen and Galbraith studied writing rather than this exact AI workflow. I’m not a researcher running a controlled trial, so my evidence is the experience of seeing whether it leads me to ideas I didn’t have at the start.
This article did. I began by comparing typing with dictation. That comparison led me to mechanical friction, which led to conceptual friction, which exposed a larger problem with AI editing. The argument changed while I worked on it. That’s the kind of change I want writing to produce.
Let AI handle the sentences
AI can make writing far less painful, and I think we should use it for that. Theres no prize for spending an afternoon fixing commas or moving the same paragraph six times.
But a finished document is only one product of writing. The other product is the change in the writer’s mind. A tool that improves the document while preventing that change has solved only half the problem.
The process I want uses AI to strip away mechanics and leave me with more room to think. It asks me to review the ideas and arrange the concepts. Once I’ve done that work, AI can help me say it clearly.
Perhaps this still counts as writing. I think it does. The writer remains responsible for the part that matters most: deciding what the work means.