Writing with AI: Does the Brush Create the Painting?
If you have read my work before, you have probably read something that AI helped me write. I tell you when it has. What I haven’t told you is what “helped me write” actually means.
I have written about this before, most directly in “I Used AI to Help Write This Piece. Judge the Output.” My argument there was that asking whether someone used AI tells us almost nothing. The better question is how they used it. Consider this the longer, more personal answer to that question: the how of my own writing process.
You might imagine that I type a prompt asking ChatGPT to write an article, wait a few seconds, and paste the result into Substack. I have tried something like that. The result was garbage.
It wasn’t incoherent. That was almost worse. It had paragraphs, transitions, and the shape of an argument. It looked like an article. What it lacked was a reason to exist. There was no curiosity pulling it forward, no process of discovery underneath it, and no person making deliberate choices about what was worth saying. The prose was performing thought rather than doing it.
Writing well with AI is much more involved—and much more human—than that. (Did you like the em dashes? I left them in for you.)
When I use AI, I begin with something I want to say or something I want to understand. From there, it becomes a research assistant, a sounding board, an editor, a skeptic, and a writing partner. It introduces ideas I had not considered and asks questions that make me examine what I think I know. Because I ask it to.
It has also changed something more personal about writing for me.
I think I am a good writer without AI, but I am a slow one. I have always struggled with freewriting because my internal editor arrives before the thought has fully entered the room. I stop over the wording, correct the punctuation, rearrange the sentence, and then edit it again. By the time the sentence is polished, I may have lost the thought that was supposed to follow it.
I don’t do that nearly as much anymore. I type quickly without worrying about spelling or punctuation, trusting that we can organize the material later. Sometimes I take ChatGPT on a walk and talk through an idea as it develops. I contradict myself, wander down the wrong path, circle back, and say things badly. The thought is allowed to arrive before I decide what clothes it should wear.
Now there is a clearer line between the flow and the edit. The cleanup is easy. Recovering a thought I edited out before I ever expressed it is not. AI has given me permission to separate exploration from polish, and that freedom has made writing more fluid. It has also made my thinking more expansive. I can follow an idea before deciding whether it belongs. I can see what is there before I begin cutting it into shape.
The stranger thing is that this practice has begun to change me. My thoughts become coherent sooner. I see the structure of an argument earlier than I used to. The tool is not only helping me organize the writing after the fact; something about the repeated process of talking, questioning, arranging, and rearranging is teaching my mind to organize itself differently. I find that fascinating.
It starts with something unsettled
When I write with AI, I begin with one of two things: a statement I want to test or a question I want to explore.
The statement is not necessarily the thesis of the eventual piece. It is closer to a proposition: something I suspect might be true but have not earned the right to claim yet. The question is more open, although questions carry assumptions too. In both cases, I begin with uncertainty.
Then I research.
There are usually several rounds. One answer reveals another question, which opens a part of the subject I had not considered. I ask for sources, follow the trail (AKA read them), and ask where the evidence disagrees (and read those too). I almost always ask some version of: How might this research be biased? What am I missing? Whose perspective is absent? What evidence would challenge the direction I seem to be heading?
That matters because AI can be too agreeable. If I ask it only to support a position, it can usually find a way to do that. It can also give me a statistic from the wrong year or a confident summary of a source that says something more qualified when I actually read it. AI research is a beginning, not a finished search. The responsibility to cite and verify does not disappear because the search happened faster.
I do not want a writing partner that simply tells me my idea is brilliant. I want one I can ask to keep me honest.
The initial research gives me material, but it usually does not give me the piece. For that, I have to explore.
Then comes the conversation
This is the part of the process I initially had trouble naming. I thought of it as conversation, which sounded different from writing. Then I thought of it as freewriting, which wasn’t quite right either.
It is both.
Sometimes the conversation is typed. Sometimes it happens aloud while I walk. ChatGPT is better for that at the moment. I bring in what I found during the research and begin turning it over. I say what seems important, what bothers me, where the evidence does not fit together, and what I think might be underneath it all. AI reflects the argument back, and I tell it where the reflection is distorted.
No, that isn’t what I mean.
That conclusion is too neat.
You are agreeing with me too quickly.
What happens if we look at it from the other side?
The exchange is exploratory in the way freewriting is supposed to be exploratory, but it answers back. I am not alone on the page waiting for my own next thought to appear. There is something on the other side of the sentence, offering a connection, asking a question, or occasionally misunderstanding me in a way that clarifies what I meant.
The misunderstanding can be useful. So can the disagreement, when I ask for it. Even the sycophancy tells me something, because pushing against an answer that is too flattering forces me to state the argument more precisely.
This conversation almost always leads back to more research. A claim that sounded plausible needs evidence. A missing perspective opens another line of inquiry. A source complicates the thing I thought I believed, so I return to the conversation with a different set of facts. Research feeds exploration; exploration reveals what else needs to be researched. The movement is circular, or perhaps spiral: I return to the same question, but each time with a little more context.
This is the real discovery phase. The thesis comes later.
The thesis is something I arrive at
Eventually, after moving between research and conversation for a while, I land on the angle. I begin to see not only what I have found, but what I want to say about it.
That is when I write the thesis.
If I began with a statement, the final thesis may resemble it, although it is usually narrower, more qualified, or pointed in a slightly different direction. Sometimes I discover that the original statement was only partly right. Sometimes I realize it was not what I wanted to say at all.
If I began with a question, the thesis is the answer I can now defend—or at least the most honest answer I can offer with the information I have.
Research gives me information. Exploration lets me find the pattern. The thesis is the claim I am willing to make about what that pattern means.
Only then do I build the outline.
If the conversation has already become extensive, I may ask AI to derive an outline from what we discussed. Other times, I give it the broad swaths of what I want to cover and ask it to arrange them. Either way, I almost always change what comes back.
I move sections. I remove ideas that are interesting but non-sequiturs. I notice when an essential point is missing or that I am asking the reader to accept something before I have earned it. I check that each section is doing work on behalf of the thesis rather than merely sitting nearby looking relevant. Often those non-sequiturs become their own posts.
The outline is not the authority. It is a proposal.
At this point I also almost always write the title myself. The title of this piece is mine. That matters because a title is not just packaging; it is the most compressed expression of what I think the piece is about. If I cannot name the piece, I may not yet understand what I am trying to say.
Drafting is still discovery
Once I am satisfied with the outline, I usually ask AI to generate the text one section at a time. Before it drafts, I explain what the section needs to accomplish, which parts of the research belong there, and how it should connect to the larger argument. Then I respond to what it gives me.
I cut sentences, rewrite passages, add context, and tell it when the language sounds generic or when it has made the argument cleaner than the truth allows. Sometimes it finds a way to express something I had been struggling to articulate. Sometimes it produces a polished paragraph that is not actually true to what I think.
A well-formed sentence can still be the wrong sentence.
The exploration does not stop because drafting has begun. I may reach section four and realize that section two needs to change. A new connection may send me back to a source, and that source may alter the thesis again. The outline bends. Earlier paragraphs move. Something I thought was central becomes a footnote, while an idea I nearly discarded becomes the heart of the piece.
That is not a failure of the process. It is the process. Writing has always been a way of discovering what we think, and AI should not remove that discovery. Used well, it gives me something to think against throughout it.
The draft grows through a continuous exchange among research, conversation, structure, and prose. The stages are real, but their borders are porous. Research changes the argument; the argument changes the outline; the draft reveals a gap; the gap sends me back into research. Each pass leaves the piece a little different, and if the process is working, it leaves me a little different too.
Then I become the reader
After the complete piece is drafted, I read it several times. I look for the places where it says something adjacent to what I mean rather than what I mean. I look for repetition, missing transitions, and claims that have grown stronger than the evidence supporting them. I ask whether every section contributes to the thesis and whether the piece holds together as a whole.
After I have done my own reading, I often ask AI to roast the draft. I discovered this while building a pitch deck and wrote about it in “Roast This Post.” The feedback had been helpful but a little too polite, so I typed, “roast this slide.” Suddenly the model stopped nudging and started telling me what was repetitive, vague, overworked, or simply not landing.
The word roast does not make the model smarter, but it changes the role I am asking it to play. Where did I bury the point? Which paragraph is doing no work? Where am I using polished language to hide an unfinished thought? I do not accept every criticism. Sometimes I argue. But that is part of the value: the exchange makes me defend the choices that deserve to stay and reconsider the ones that don’t.
Sometimes I put the draft into a different model for another review. If Codex helped me write it, I may ask Claude to critique it; if Claude helped with the draft, I may bring it to Codex. Usually Claude is the reviewer. (I have not found Gemini nearly as useful for writing or reviewing, although other people may have a different experience.)
A second model is like bringing in an editor who was not in the room for all the earlier conversations. It has not grown accustomed to the draft’s assumptions, so it may notice something the first model and I have stopped seeing. I ask it to identify the weakest section, tell me what a skeptical reader would challenge, and point out where the argument sounds more certain than the evidence allows.
Sometimes the piece goes through three or four drafts. When the argument finally feels settled, I ask AI to copy edit it. Then I read it again, make more changes, and paste the finished version into Substack.
You may be noticing something familiar here. Research, conversation, drafting, revision, feedback, more revision. This is not entirely unlike the age-old process of writing.
You are right.
Writers have always talked through ideas, worked with editors, reorganized arguments, rewritten paragraphs, and discovered halfway through that the piece they thought they were writing was not the piece they needed to write. AI has not eliminated that process for me. It has made the process more conversational, more available, and sometimes more rigorous. I have access to a partner I can question whenever the thought begins to move.
The important part is that I remain in the room.
So, what is the point, you may ask. First, it’s faster. It takes me about a quarter of the time as it would have to go through this process. Second, it enhances my writing. It keeps me honest, and brings in points I hadn’t considered. And third, it’s pleasurable. It just is. I’ve always loved to learn, and using AI as a writing partner makes learning so damn fun. That’s why I do it.
I am the human and the driver
Does this process sound like cognitive offloading to you? Because it actually feels more rigorous than the writing I did in college. It is my question or proposition that starts the process. I direct the research. I decide which evidence matters and what the evidence can reasonably support. I choose the argument, shape the structure, write the title, and determine what stays.
AI contributes ideas, connections, criticism, structure, and language. Sometimes it formulates a sentence better than I would have. I have the humility to say that its contribution is real.
But I accept or reject thecontribution. I decide what I believe. I put my name on the piece, which means I am responsible when it is right and responsible when it is wrong.
That is why the complete-article prompt produced garbage. I had temporarily removed the part of the process that gave the piece purpose: myself.
What does “90 percent AI” mean?
I am transparent about my use of AI. I do not use it to write every piece, but when I do, I select the disclosure that says AI helped me write it.
Substack also has a feature called Scan for AI text, which uses a tool called Pangram to estimate what percentage of a piece is human-written or AI-assisted. My pieces often come back as more than 90 percent AI.
I am not entirely sure that is how I would assess them.
Did AI formulate many of the final words? Yes. If I paste the copy-edited version from the canvas into Substack, I understand why a detector sees its fingerprints everywhere. At the sentence level, maybe the percentage makes sense.
But the number cannot tell you who formed the question, directed the research, asked about bias, noticed what was missing, rejected a weak argument, or changed their mind. It cannot tell you who decided the piece was worth writing. A detector can examine patterns in the finished prose; it cannot reconstruct the intellectual history underneath it.
That does not mean AI made no meaningful contribution. It means authorship is more complicated than counting who formulated each word.
I am being changed by it
The most interesting part of this process is not that AI helps me produce a polished article more quickly. It is that working this way is changing how I think.
Because I regularly ask AI to test my assumptions, I have become more likely to look for them myself. Because we organize arguments together, I see their structure sooner. Because I repeatedly ask what is missing, I am more conscious of absent perspectives before I reach the end.
The conversation is slowly moving inside my head. I begin anticipating the questions before I ask them.
Better organization in my writing has brought better organization to my thoughts, which is why I do not experience this as offloading my brain. I experience it as enhancing it. The distinction matters. Offloading would mean asking the machine to think so that I do not have to. This process asks me to keep thinking: to respond, evaluate, resist, choose, and sometimes begin again.
That does not make the influence automatically good. AI can make a weak argument gleam. It can smooth away the strange and distinctive parts of a voice. It can reinforce the premise of a question instead of challenging it, and it can be confidently wrong. Writing well with AI requires an awareness of those tendencies and a willingness to push back against them.
The goal is not to surrender judgment. The goal is to exercise more of it.
Does the brush create the painting?
AI helped me write this piece. I brought the idea and described the process. AI reflected it back and gave it an initial shape. I corrected the pathway when the shape was wrong. I asked for a different rhythm when the prose did not sound like me. I wrote the title. We made an outline together, and I changed it. AI drafted sentences; I read them, questioned them, cut them, and changed the direction.
So who wrote it?
If writing means physically formulating every sentence, then AI wrote a great deal of it. If writing means choosing the question, pursuing the answer, deciding what the evidence means, shaping the argument, and taking responsibility for the result, then I wrote it.
The honest answer is that we wrote it together, but collaboration doesn’t erase authorship. It asks us to think more carefully about where authorship lives.
Does the brush create the painting?
Anthralytic is a strategy and evaluation studio for mission-driven organizations. We help teams use data, evidence, and AI to understand what is working, make better decisions, and build systems that serve people—not the other way around.
