Assisted, not delegated: writing
The space between human and AI
A piece came across my feed last week from someone in my extended orbit. Good premise, the kind of historical thread that promised a real essay if it got the right hands. The caption underneath was confident: two months of extensive reading and research, the author’s proudest piece of work, history is truly humbling. I opened it expecting to learn something.
I had seen this shape before. A single rhetorical move repeated for three thousand words. Couplets, antitheses, the same syntactic gesture wearing different costumes. The specific material a serious reader of the subject would expect never quite arrived; the named references could have been gathered in a single afternoon on two well-known forums. The piece was atmospheric where it should have been textured, abstract where it should have been concrete. The kind of writing a language model produces when asked to sound essayistic about a subject the prompter has not actually studied.
I tried to raise it with a mutual friend who knows the author better than I do. The reply came back inside a minute: I’m not gonna say that to the guy. The substance never got litigated. The conversation closed before it opened. And I sat with that for a while, because the friend was not wrong by any rule of friendship I would want to defend, and the piece was not going to improve on its own, and somewhere in the gap between those two facts was the thing I have been circling for months without writing down.
I have been circling this for months without writing it down. The first piece in a series, because the problem extends past writing, and the answer in each domain is shaped a little differently. Writing is where it starts, and writing is where I have done the work, so writing comes first.
Earlier this year I published a piece called Care as a Discipline, about how painting and celestial navigation survived their automating tools. The argument was that a craft does not die when a machine arrives to do its job; it survives when its judgment can be moved somewhere the machine does not reach. Painting found higher ground in interpretation, in deciding which appearance was worth holding. Navigation found lower ground in resilience, in being the position you can still recover when the convenient signal has been jammed or spoofed. Ground existing, the piece argued, was not the same as anyone standing on it. The discipline of care was the early decision to move the craft’s judgment while moving it was still cheap, before the loss had happened and the rebuild was forced.
I wrote that essay in the abstract. The crafts in it were historical. Painting and navigation were the analogies, and the verdict from 1839 was the metaphor for the verdict being delivered now against a long list of crafts at once. I did not, in that piece, take any specific craft and stand on its ground in public, with my own name on the work. Writing is the craft I am trying it with first. AI is the tool. The question is what the higher ground looks like in practice for a writer who refuses to either abandon the tool or surrender to it. The distinction between assisted and delegated is the answer I want to offer, and the reason it matters even though almost no one in the current public conversation can see it.
The public conversation about AI and authorship has settled, for now, into two positions that talk past each other.
The first position belongs to writers like Andrea Bartz, the novelist who has spent the last year arguing on her Substack that generative AI is eroding the trust between readers, writers, and publishers. Her position is principled, technically informed, and absolute: any contact between a language model and a published text contaminates the text. The detection regime is failing (false positives are ruining careers, certifications can be bought for ten dollars and ninety seconds of work), and her response is to call for harder infrastructure. Human Authored certifications with legal teeth. Behavioral signal capture during composition. Cryptographic proof of authorship. A writers’ union with the muscle of the WGA. The whole institutional apparatus, built from scratch, to defend a category that used to defend itself.
The second position belongs to writers like Jo Shaw, who acknowledges the AI tells (rule of three, “not this but that,” em-dashes) and immediately deflates them: humans invented all of these patterns first, and it is entirely possible that writing which screams AI is just bad human prose. Shaw’s anxiety is about the witch hunt. The Mia Ballard case sits in her essay as a warning: an author whose career was destroyed on detection-based suspicion, who may or may not have done what she was accused of, and whom we will probably never be able to clear or convict. Shaw closes by asking her readers whether a Human Authored certification would help, or whether the standard should be innocent until proven guilty.
Both positions are responding to something real. The trust erosion Bartz describes is happening. The witch hunt Shaw fears is happening. The Ron Charles experiment Bartz quotes at the end of her piece (using ChatGPT to generate five hundred words from an interview transcript, paying ten dollars, receiving a certificate of human authorship in ninety seconds) is evidence the verification regime cannot bear the weight being put on it.
Both positions treat the field as binary. Human-written, or AI-written. Trustworthy, or not. In Bartz’s framework these are the two categories, and the work is to defend the first from the second. In Shaw’s framework these are still the two categories, and the work is to keep accusation from running ahead of evidence. Neither writer can see the practitioner I am trying to describe, because that practitioner doesn’t fit in either box.
There are at least three practices currently being squeezed into two categories, and the collapse is where most of the harm comes from.
The first is unassisted writing: a human at a keyboard, no model in the loop, the entire judgment chain in one head. This is what Bartz defends and what Shaw worries is being falsely accused. It is the older practice, and it is not going anywhere, and it deserves the protection both writers are trying to give it.
The third practice (I will come to the second in a moment) is delegated writing. A human with a topic and a vibe prompts a model to produce content. The model generates. The human pastes, posts, signs. The judgment chain has one decision in it: post. Sometimes the human adds a caption claiming research, expertise, pride. The signature is a forgery in a specific sense: it promises a presence the text cannot deliver. The Ron Charles experiment is the pure form of this. The piece that came across my feed last week is a less pure but recognizable example.
Between unassisted and delegated, there is a middle practice that the public conversation does not currently see, and that is the practice I want to defend. Assisted writing. A human with something to say uses a model as part of a process the human controls. The thinking, structuring, selecting, weighing, cutting: those remain the human’s. The model is a tool inside a discipline. The output is signed by the human because the human is in fact present in the output. The judgment chain has hundreds of decisions in it, most of which the reader will never see, all of which leave traces in the text if you know how to look.
The collapse of these three into two is not a neutral simplification. It hurts everyone. It hurts unassisted writers, who become paranoid and start defending their humanity against accusations that should never have been levelled. It hurts assisted writers, who become invisible, their disciplined practice read as either suspicious (Bartz’s view) or naive (Shaw’s view). And, most consequentially, it lets delegated writers hide inside the assisted category by adopting the same disclosure language. Yes, I used AI. I’m transparent about it. Everyone does. The honest practitioner and the dishonest one say the same sentence, and the sentence stops carrying information.
This is the bucket problem. When “AI-assisted” comes to mean anything that a model has touched, including pieces no human actually thought, the term ceases to function as a signal. The honest writer pays the bill for the dishonest one. The bad money drives out the good until the category itself collapses.
Bartz writes about authors who might be tempted to publish AI-assisted work as human-written, and predicts what they will tell themselves: I totally rewrote what it generated, I was just using it to get the juices flowing. Her implication is clear. Anyone who explains their AI use this way is rationalizing. The honesty is performed, the discipline is fiction, the byline is still a forgery.
She is partly right. Some people do say those sentences as cover for delegation. The cover is real and the sentences are sometimes lies. If I am going to argue for the assisted category, I owe the reader more than my insistence that I am the honest kind.
Most of the disclosure language Bartz is suspicious of is unverifiable by design. The writer says I rewrote it and the reader has no way to check. I used it to get the juices flowing and the reader has no way to check. The phrase is offered as a credential, and the credential carries no proof. Bartz is right to find this hollow.
What I can offer in place is a record. I draft in markdown files, in a git repository, with commits at meaningful intervals. The commit messages describe what changed and why. The version history shows the order of decisions: which paragraph existed in the first draft and survived, which one came in on the fourth revision after a structural rethink, which sentence I argued with myself about for three commits before cutting it. The conversations I have with the model are saved as transcripts. The frameworks I apply to revision (the Humanization Framework, now in its ninth version, the Cognitive Framework for AI-Assisted Intellectual Work, the voice synthesis systems for the pseudonymous projects) are public documents. Anyone who wants to know what assisted writing looks like in my practice can read the methodology and inspect the trail.
Bartz herself, in the same essay, points at the only solution the current moment actually allows. She quotes Vera Kurian: the right detection unit is not the content of the text but the behavior of its production. Save your drafts. Show your revision history. Demonstrate the labor. Kurian writes about Google Docs autosave and slow incremental changes. Bartz writes about her own use of Google Docs, with a self-conscious aside about the irony of using a Google product when Gemini is part of the problem.
The irony she names cuts further than her essay registers. Writers worried about AI’s contamination of authorship are documenting their human process using infrastructure built by one of the largest producers of generative AI in the world. The contradiction is structural. Writers who want to defend human authorship reach for tools built by the companies they are defending against, because nothing else in the consumer software stack does what they need.
The software industry has been arguing about authorship of intellectual work for thirty years, and it produced a tool for exactly this problem. Git, and the public hosting platforms built on top of it, were designed to make authorial decisions visible: who changed which line, when, in what order, with what stated reason. Every commit is a timestamped claim. Every diff is a record of judgment. The branching, merging, and review patterns are a documented social process for how decisions get made and by whom. A writer who drafts in git can produce, by default, exactly the kind of behavioral evidence Kurian is asking for, and can do it on infrastructure that has no entanglement with any AI company’s training data ambitions.
The fact that this is unusual in the writing world says less about writers than about how the two communities have developed in isolation from each other. Coders have an authorship discipline because their work has been treated as intellectual property under legal pressure for decades. Writers have an authorship discipline because their work was never automatable enough to require explicit defense. Now that it is, the writers are reaching for the coders’ tools. They should. Those tools work.
I should say, before I go further, that I started where the delegated writers started.
The first pieces I wrote with model assistance were mostly the model’s. I prompted, it generated, I tidied, I posted. The output had the same shape as the piece I described at the top of this essay, and if I had kept going that way I would now be producing the same hollow atmospheric work in a different domain. I had the same initial reaction everyone has. The model produces fluent text fast. The fluent text looks like the writing I would do if I were better. The combination is intoxicating, and the intoxication is the problem.
What pulled me back was reading my own pieces a few weeks after publishing them and recognizing that they did not contain me. The voice was a flattened average of the voices the model had been trained on. The arguments were the ones the model could reach without effort, which meant the ones that had been made many times before. The reader’s attention I had asked for was being repaid with content I had not actually generated. The signature was the forgery Bartz describes.
The work since then has been the work of climbing back into the chair. The Humanization Framework began as a private discipline for catching the specific places where my prose had collapsed into model defaults: the em-dashes, the binary corrections, the staccato lists, the topic-sentence orthodoxy, the prestige vocabulary, the synonym rotation, the period overuse. It is now a forty-nine-rule document across twenty-three sections, applied at the revision pass, mode-selected by genre. The Cognitive Framework came later, as a companion: a structured process for the epistemic side of the work, ontology mapping and adversarial audits and the things a model will not do for you because they are the parts where the thinking happens. The voice synthesis frameworks are for the cases where the voice on the page needs to be specifically not mine (pseudonymous projects, particular registers), and where the discipline is correspondingly harder.
None of these frameworks make a model write better. They make me write better when a model is in the loop. They are the answer to the question what is the human supposed to be doing while the model is fast. They formalize the judgment work that distinguishes assisted from delegated, and they do it in public, in version-controlled documents, so that anyone who wants to inspect the practice can. They are, in the language of the earlier essay, the way I have tried to move my own judgment to higher ground while the move is still cheap.
I have not arrived. The frameworks are imperfect, the discipline is uneven, some pieces come out better than others. But the direction is the one Bartz is asking for, and the methodology is more rigorous than the Google Docs autosave she settles for, and the work happens inside a category, assisted and not delegated, that her essay cannot quite see.
Back to the friend.
The reply was I’m not gonna say that to the guy. No counter-argument about the piece, no defense of its quality, no claim that I had read it wrong. Just the refusal. And the refusal was reasonable. Friendship is not editorial. Unsolicited criticism of someone’s creative work lands as an attack on identity, not on output, and most people will not do that to a friend. I would not have wanted him to either, if I had been thinking about it purely as a question of friendship.
But the piece was not a private journal entry. It was published with a caption claiming two months of research, the author’s proudest work to date, history is humbling. The author was making public claims and asking strangers to read his work as expertise. Once a piece enters public space with substantive claims, it is in a different conversation, and the friendship-protective instinct stops fitting cleanly. The author was asking the internet to take him seriously.
The reason this matters for the larger argument is that the writer I read last week, and the thousand writers like him quietly posting delegated content to small Substack audiences every day, sits in a place where none of the institutional remedies Bartz and Shaw are debating will ever reach. He has no publisher to cancel him. He has no editor to push back. He has no contract that requires originality. He is not famous enough to be falsely accused in a witch hunt, and he is not famous enough to be defended by a union. He is below the threshold where Bartz’s certification regime applies and above the threshold where Shaw’s detection anxiety attaches. He is just a person posting, and the only people positioned to register that the work has thinned out are the people who know him personally.
Those people will not say anything. Not because they cannot see what I saw. The mutual friend almost certainly sees it too. He has read the piece, registered the shape, formed the judgment, and kept it to himself, because the social cost of voicing it is real and the social reward is zero. This is how the critical reasoning of an entire community becomes invisible. Every individual reader privately notices the hollowness. No reader says it out loud. The published record looks like consensus approval. The actual distribution of opinion is silent.
The remedies Bartz and Shaw propose are upstream, at the level of publishing infrastructure, professional certification, and detection technology. Those remedies might eventually defend the big-publisher novel from the AI-delegated competitor. They will not touch the small-platform writer who has decided that two months of prompting counts as research and his name on the post counts as authorship. That writer will never face Bartz’s union, never need Shaw’s certification, never trigger a detection sweep. The only correction he could receive would come from the people closest to him, and the people closest to him have already calculated that the correction is not worth the friendship.
The friend should not necessarily have spoken. But his silence is the actual mechanism by which delegated writing proliferates at the small-platform level, and no amount of institutional engineering will substitute for it. If we want the assisted category to survive, the work cannot only happen in policy debates and certification schemes. Some of it has to happen at the level where I started this essay, with one person noticing that another person’s published work does not contain them, and deciding what, if anything, is appropriate to say.
I do not have a clean answer to that question. I tried to raise it with a mutual friend and the conversation closed before it opened. The essay is what I have instead, and the writer who recognises himself in it is welcome to take what is useful and leave the rest. But the essay is the second-best solution. The first-best would have been a quiet conversation between two people who already trust each other, and that conversation did not happen, and it almost never does.
I write under my own name. I disclose my use of models openly. I publish the frameworks. I draft in git. I sign work when the work bears my judgment, and I withhold the signature when it does not. The byline I put on a piece is a promise that the thinking behind the piece is mine, that I have weighed and selected and revised, that the residue on the page is what I meant to leave there.
That promise depends on what I did with the model, regardless of whether a model was in the loop. Assisted, not delegated, is the line. Detection software cannot draw it, because detection looks at the output and the difference is in the process. Only the practitioner can draw it, and the practitioner has to be willing to show their work.
The question that follows me from the piece I read last week is whether the rest of us (the ones doing the slower, harder, assisted work) will hold the line on what the signature means, or whether we will let the category collapse and lose the distinction altogether. People will go on posting model output under their names, and the social cost will eventually catch up with them, or it will not. That is a separate question.
I am writing this series because I think the line is worth holding, and because the only way to hold it is to make the practice visible enough that the difference can be seen. Care as a Discipline set out the frame in the abstract. Writing is the first domain. Coding is the second: the tool I use to defend my writing was built by coders to defend their code. Creativity in general is the third, and by the time we get there I hope the categories will have done some of the work for us.
For now: assisted, not delegated. The signature still means something. Some of us are still in the chair.
References
Andrea Bartz, “So…how can you prove your work’s not AI?” Substack, 2025.
Jo Shaw, “Moments of Being #2,” Thought Couture on Substack, 2025. (See the section on AI witch-hunting.)
Marco Brondani, “Care as a Discipline: How painting and celestial navigation survived automation.” marcobrondani.com, May 2026. https://www.marcobrondani.com/p/care-as-a-discipline