The front page of synthesiswriting.org states the position in two short sentences: the writer writes, the AI assists. The first sentence is easy to agree with. The second is where a fair reader should push, because “assists” has been stretched in public use until it covers nearly everything. A spell-check is called assistance. So is handing a model your notes and publishing what comes back under your name. So is a model trained on your archive turning out columns in your voice. A word that covers all three is not drawing a line. It is hiding one.
The natural question is a percentage question: how much AI is acceptable? Percentages measure the wrong thing. A spell-checker touches every word I publish and takes nothing from me. A model that answers the question what do you think about this has touched almost none of the words and taken the only thing that mattered.
Provenance, not percentage. Three questions decide it: whose experience is on the page, whose position does the piece take, and whose voice carries it. Assistance is whatever help leaves all three answers unchanged. The moment help changes an answer, it has stopped being assistance, whatever it is called.
The third question is the one that needs defining before it can do any work, because “voice” is used two ways and they lead opposite directions. If voice means who first put those particular words in that particular order, then any drafting help fails the test, and so does a copy editor who rewrites your paragraph. If voice means the expression I deliberately adopted and will defend as how I sound, then the test is about what you kept and why, and it can survive help that the first definition rules out. I use the second, and I will argue for it below rather than smuggle it in. What the second definition cannot survive is drift you did not notice, which is why most of my machinery points at exactly that.
Here is the line as I hold it at my own desk, with my own name on the work.
What I ask of the model
Most of what I delegate shares one property: it operates on material that already exists. Facts in the world. Sources already published. Sentences I have already written.
Research is the largest share. Find the study I half-remember. Find the writer who made this argument long before me, so I can credit them. Find the passage in a two-hour interview where the source finally said the thing.
Verification is the share I would defend hardest. Check every factual claim in this draft and show me which ones are wrong. I want the list, not reassurance. A model that catches my misremembered date before a reader does has served the writing without writing any of it.
Structural critique: tell me where the argument breaks, which paragraph a hostile reader would quote against me, what I asserted without earning. The most useful reviewer is the one arguing back.
Pattern detection: scan my draft for the tells of machine prose and for emptiness, and flag both. More on that below.
And the mechanics, settled long ago: spelling, grammar, reformatting a piece to a different length for a different venue. Nobody’s authorship was ever threatened by a spell-checker.
In every one of these the model retrieves, checks, or challenges. Those are the jobs of a research assistant, a fact-checker, and an editor whose role is to argue with me. Honorable jobs, all three. None of them is the writer.
And then there is the one that makes people uncomfortable, including me, which is why an essay drawing this line is worthless if it skips past it. Sometimes I have a model draft.
Not from a standing start. What I hand over is a position I have already formed, an experience I actually had, and an argument whose shape I have decided. What comes back is a rendering of that. Then it gets taken apart, because a first rendering is reliably fluent and reliably a little wrong: it smooths the part that should stay jagged, it reaches for the phrasing anyone would reach for, and it states my position slightly more confidently than I hold it. Nothing stays that I have not decided to keep.
I want to be exact about why this is not the thing the byline forbids, because the difference is one step and the step is easy to skip.
What stays mine
The position. What a piece argues is something I formed and will defend in public, which is different in kind from a position assembled on demand out of the internet’s median opinions.
The lived material. The model was not in the rooms I write about. It cannot supply the meeting, the mistake, the conversation, the years of watching an industry change from the inside. When a detail appears in my work as experience, it happened to me. There is no assisted version of that.
The thinking that happens in drafting. This is the one people mean when they say the sentences, and it is worth separating from the typing, because they come apart. Drafting is where I find out what I actually think, and what I find is regularly narrower and less flattering than what I sat down believing. A model will gladly carry me past that step, straight to fluent paragraphs about what someone in my position would probably think. That is the theft, and I would be the one consenting to it.
So the rule I hold is about sequence, not about who types. I do not hand over a subject and take back a position. By the time a model renders anything under my name, the finding-out has already happened. If I ever cannot say what I think before I ask, that is the signal to stop asking and go work until I can.
And the judgment. An assistant can verify that a claim is accurate. It cannot decide that the claim is worth a reader’s time. Deciding what is true and worth saying is the writer’s job, and it does not delegate.
Adversarial review, not ghostwriting
The mechanics of my practice are public, so I will name them plainly. I keep catalogs of the patterns that mark machine-generated prose and the patterns that mark empty prose whoever produced it. They ship as open-source instruction modules a writer’s AI tools can load: one for AI-generation patterns and substance tests, one for the craft principles a review should push toward, one for the older human failures that predate AI entirely. Slopcheck, a free checking tool I built on those catalogs, runs the same tests for anyone.
My own drafts go through that tooling before they go anywhere else. Not because I suspect myself of pasting in machine prose. Because I know what a model in the loop does to a writer’s ear. The article that introduced synthesis writing exists because I caught a sentence in my own draft that sounded right and belonged to no one. The drift toward the median voice is gradual, it flatters you the whole way down, and I do not trust myself to hear it unaided. So the model’s standing assignment on my writing is adversarial: catch the borrowed phrasing, fail the paragraph any writer could have produced, run the deletion test (if a passage can be cut without losing a claim, it goes) against everything that merely sounds finished.
Positioned as a critic, the model makes the writing more mine, because everything that survives the review is something I chose to keep. That is also the honest answer to the obvious objection about the paragraphs above: if a model can render my sentences, what stops the rendering from quietly becoming the writing? The review does, and only because it is aimed at exactly that. The tests I run hardest are the ones that fail a paragraph for sounding like anyone. A draft that came back fluent and survived a pass built to punish fluency has earned more scrutiny than a draft I typed myself, not less.
Positioned as the author of the position rather than the prose, the model makes the writing less mine with every accepted suggestion, and by then no amount of editing recovers it. Same model. Same desk. Opposite crafts.
How the line fails
It fails in both directions.
Crossed, it produces prose nobody stands behind. The test is not whether a model touched the sentences. It is what happens when you push on the piece: ask its author to defend a paragraph, extend the argument, say what they would concede. Where the position was formed before the drafting, there is something underneath, and the author can go further than the page did. Where it was assembled on demand, there is nothing to push on, because the position was never held, only rendered. Much of the slop that has flooded the internet these past few years is that second thing at industrial volume, but it does not require a spam operation. It happens one respectable byline at a time.
Held too tightly, the line fails quietly. If I refuse to let a model check my facts, hunt down the source I misremembered, or flag the paragraph any publication could have run, I publish errors and slack that a review pass would have caught. My authorship was never at risk from a fact-check. Purity that refuses verification is not protecting the craft. It is protecting the mistakes.
One distinction, so this is not misread: the writer who declines these tools entirely is practicing the craft exactly as legitimately as I am. A calligrapher is not failing at writing by refusing the word processor. The waste I am naming belongs only to those of us who have already chosen to work with the tools and then keep them away from the work they do best. Refusing the whole practice is a position. Refusing the fact-check inside the practice is just a worse draft.
The test for your next draft
Before the next piece goes out under your name, run the three questions over it. Whose experience is on the page? Whose position does it take? Whose voice says it?
Help that leaves every answer “mine” is assistance, and I take all of it I can get: the fact-check, the found source, the structural argument, the adversarial read, and yes, on plenty of days, the rendering of a position I had already worked out. The first time an answer is not “mine,” no percentage will rescue the piece, because the thing the byline promises is no longer in it.
If that last admission makes you trust this essay less, I would rather have it than the version where I left it out.
That is where assistance ends. Not at a word count. At the point where the writing would stop being yours.