What are we trying to protect when we label AI?
The provenance of the sentences is not the same thing as the provenance of the thought.
This essay began as thirty pages of my own exploration and a shorter draft. I developed the published version in conversation with an AI collaborator, using that exchange to clarify the structure and language. The experience is mine, the ideas are mine, and I remain responsible for every sentence. The finished expression was made together.
Like many people, I watched the response to Substack’s new AI-detection feature and initially thought, “Yeah, okay. For people who are here to practise the art of prose (which is a significant part of Substack’s original culture) I understand why this matters.”
And then I thought, “Wait. That’s not what I’m here for though..?” Does that mean I don’t quite belong in the Substack community, with or without AI? Bear with me, because I think this question applies to much more than just Substack.
I started writing here because I wanted to participate in conversations about the things I don’t understand in the world, as well as the things I may be beginning to understand.
I want to think with other people and poke at what we don’t know. I want to explore questions across the many languages of psychology, history, sociology, science, technology and the experience of being human, often all at the same time.. this is tricky work.
I tend to think in multiple dimensions/perspectives at once, and putting language around that can be challenging. Before I try to write an “essay,” I may spend hours exploring an idea, drawing on the lifetime of experience behind it and following each thread far enough to see how it connects with the others. Left to my own devices, I can easily produce thirty pages.
Those thirty pages are useful to me - they help me separate the threads before weaving them back into some kind of coherent understanding. They are part of how I think. But several decades of working with other people have also taught me that translation is a critical part of thinking together.
The thirty-page exploration that is illuminating for me may be almost impenetrable to someone else. Other people may still be interested in the questions I’m asking, but I don’t naturally create many on-ramps for them to enter the thinking space.
And that leaves me with a dilemma. Do I keep the exploration largely to myself because I’m not especially interested in developing the craft of literary writing? Do I spend hours turning every thought into carefully honed prose, even though that isn’t the work that energizes me? Or can I collaborate with something or someone that helps me translate the complexity into a form other people can enter?
For me, Substack is primarily a community of thinkers. I’m not here to compare my prose with the work of poets, novelists and essayists for whom language itself is the artistic medium. I love consuming that kind of writing, but it isn’t the practice I’m pursuing.
That doesn’t make the language irrelevant. If I want to think with other people, I need to communicate clearly enough for them to find their way into the thought. But for me, language is in service to the inquiry. It is not the central object being created. I think this may be one reason the conversation about AI and writing feels more complicated than the labels allow.
People who say they can recognize when AI has written something are often responding to something real. There is a certain texture to generic, machine-produced prose, and there is something unmistakably human about art that translates a particular experience through the choices, judgment and sensibility of an artist.
I don’t want that distinction to disappear, and I also don’t think everything we publish is trying to be that kind of art. Sometimes the purpose of writing is to make thought available and to create a meeting place where an idea can be examined, challenged and expanded by other people. That’s usually what I am doing here.
My process tends to look something like this:
I explore. Sometimes at considerable length. Then I write a shorter narrative that I hope will create a conversation.
After that, I may share both the exploration and the shorter piece with AI. I ask where I have inadvertently erased other perspectives. I ask where I have created another off-ramp through verbosity or complexity. I ask where the logic doesn’t hold together, or where a reader might misunderstand the point I’m reaching for. And, yes, sometimes I ask for a cleaner version.
Then I go through the suggestions. I keep what feels accurate. I reject what doesn’t. I put things back when the refinement has removed too much of me. I make the final judgment about whether the piece still says what I mean and whether I am willing to stand behind it.
Usually, the result isn’t radically different from what I wrote. But it is more accessible to other people. And this makes me happy.
Until now, I might have described this as using AI to refine my writing, but even that no longer feels quite accurate. I could refine the writing myself because I understand the changes that would make it tighter. What I am not interested in doing is spending another hour honing every essay when the part I care most about is the inquiry itself.
Or I can co-write it with AI. I can remain responsible for the experience, the questions, the underlying thought and the judgment about what is true. AI can become a partner in structure, compression, clarity and translation. It can help me notice gaps, anticipate misreadings and create the on-ramps I often fail to create on my own.
The finished piece then emerges from a relationship.
That isn’t the same as asking a machine to generate a thought and placing my name on it, but it’s also more than spellcheck. The collaboration affects the expression, and sometimes the exchange helps me recognize an implication I had not yet articulated, and I’m comfortable being transparent about that.
What makes me uncomfortable is the assumption that the most meaningful question is simply: Did AI contribute to these words?
That tells us something about the production of the sentences, but it may tell us very little about the production of the thought.
Who originated the inquiry?
Whose experience is being examined?
Who exercised judgement?
Who decided what belonged and what did not?
Who is willing to be responsible for what was finally published?
A detector cannot answer those questions. It can estimate patterns in language, but the provenance of the sentences is not the same thing as the provenance of the thought.
And this becomes even more important when we look beyond people like me. I have been reading perspectives from people who write in English as an additional language. They may have complex and beautiful thoughts to contribute but face a language barrier that makes it difficult for others to meet those thoughts. AI can help translate not only between languages, but between a person’s internal understanding and the conventions through which others can recognize it.
I don’t want to romanticize that or use someone else’s experience to justify my own. I just think it shows how many different human purposes are being compressed into the single category of “AI-assisted writing.”
Some people are making literary art, some are translating, some are organizing research, some are learning to communicate in a new language, some are using AI to manufacture plausible authority they have not earned, and some are doing what I am doing: developing their own thinking in conversation with another form of intelligence and using that collaboration to make the result more accessible. These practices are not ethically or creatively identical just because AI touched the language.
And maybe that’s what troubles me about the current discussion. A label that appears to offer transparency may actually erase the distinctions we most need to understand. I can appreciate why readers want to know what they are encountering. I want to know, too. I don’t want human experience displaced by vast quantities of frictionless content made without curiosity, accountability or care.
But if we are trying to protect authentic participation, we need a richer understanding of authorship than a percentage produced by a detector. Maybe disclosure should describe the process rather than classify the product. Tell me:
whether AI generated the premise.
whether it assembled the research.
whether it drafted the language, challenged the reasoning, translated the original or served as an editor.
what the human contributed and who stands behind the result.
That would help me understand what I am reading. “AI-assisted” doesn’t tell me very much at all.
Maybe there are different arenas here not ranked above or below one another, but organized around different purposes. There should be space for writers who want to explore human-generated language as an artistic form. There should be space for people who are thinking publicly and want help creating an opening through which others can join them. There should be space for work created through transparent collaboration between different kinds of intelligence.
We don’t need to pretend these are the same practice, but we also don’t need to turn the difference into a moral judgment. The question isn’t only whether a machine touched the language. The more interesting questions are where the thinking came from, how judgment was exercised, what kind of relationship produced the work, and who is willing to stand behind what was published.
Those are the distinctions I want our labels to help us see, and, fittingly, this essay is now an example of the thing I am describing. It began as my own exploration and my own draft. I then developed it in conversation with an AI collaborator, using that exchange to clarify the structure, examine the logic and make the language easier to enter.
The experience is mine. The questions and underlying argument are mine. I remain responsible for what is published.
The final expression was made together.
Post Script:
After co-writing this essay, I ran it through Substack’s detector and it classified the result as:
AI: 49%
AI-assisted: 19%
Human: 32%
I was shocked by that result, not because AI had made no contribution, but because the percentages bore so little relationship to my experience of creating the piece.
So I tried applying a different set of filters. Here is my own account of where the work came from:
Impetus for the essay: 100%
Lived experience behind it: 100%
Human original inquiry and ideas: 100%
Human exploratory thinking: 100%
Human first draft: 100%
Human core argument: 100%
Logic: Human 95%, AI 5% developed through dialogue
Structure of the final draft: Co-created, then human edited final draft
Final language: Substantially AI-generated, under my direction
Selection and final judgment: 100%
Human Responsibility for publication: 100%
These percentages can’t be added together. They are not competing portions of a single pie, and that’s exactly the point.
A piece of writing is not one undifferentiated substance called “content” - it contains experience, inquiry, ideas, reasoning, structure, language, judgment and responsibility. Different participants can contribute differently to each layer.
Substack’s detector appears to be evaluating the statistical characteristics of the final language. That may be one relevant piece of information, but the presentation turns that narrow measurement into something that looks like a comprehensive account of authorship: it tells the reader this essay is only 32% human.
My experience is that it is entirely grounded in human thought and experience, expressed through a genuine collaboration in which AI contributed substantially to the final language and structure.
Both accounts acknowledge AI involvement. But they tell profoundly different stories about what was created and who created it. The question isn’t which percentage makes me look better, it’s which account helps a reader understand the work.
