Did You Write This?
On AI writing, intellectual distance, and the person we came to read.
In this post
I recently published an essay about artificial intelligence and consciousness. When a friend read it, his first question wasn't about the arguments. He wanted to know whether I'd written it, or AI had.
It's a fair question. I supplied the ideas, discussed them, argued over them, and decided what I wanted to say. The prose was largely the work of a language model. I told him so. But I was frustrated. I'd hoped to discuss the arguments. Instead, we were discussing the production process.
At first, I thought his question was about effort. Writing traditionally takes longer than reading: someone spends hours organizing an idea, and the reader offers their attention in return. Spend fifteen minutes reading something that took thirty seconds to generate, and you might feel cheated.
But effort is a poor measure of an argument. A brilliant observation doesn't improve because its author agonized over it for a week. Nor can a reader tell whether an AI-assisted essay emerged from hours of discussion or a single prompt.
Now I suspect my friend's question had more to do with distance. He wanted to know who he was having a conversation with.
The essays I never wrote
I've wanted to write for most of my life. Ideas, observations, connections between unrelated subjects: the desire to express them was there. The writing usually wasn't.
Something can feel intelligible in my head, a network of associations I can navigate almost intuitively, while turning it into an orderly sequence of sentences feels grueling. Choosing words, rearranging clauses, preparing the reader for the next thought. Like toiling in the mines of my mind, excavating pieces of something I understand and assembling them into a word totem, an idea idol, that another person might recognize.
Conversation, on the other hand, is wonderful. I can begin somewhere, discover a connection, reconsider an assumption, and move toward something else. The intellectual work happens as the conversation unfolds. I can spend hours doing this without noticing the time.
Ask me to construct an essay from the resulting thoughts, and the enterprise becomes exhausting. There are other things I'd rather do with that effort: examine another idea, follow another curiosity, or attend to the ongoing maintenance of this decaying temple of flesh.
So I mostly kept my ideas to myself. AI changed that. I can explore an idea through conversation and have that conversation distilled into something another person can read. The work of thinking hasn't disappeared, but translating it into prose has become less burdensome.
I haven't stopped thinking because I can use AI to write. I've started publishing because I no longer have to do all the writing myself. The alternative wasn't necessarily an article written entirely by me.
In all likelihood, the alternative was nothing.
We have ChatGPT at home
When I read a personal blog, I'm looking for more than information. I don't read Fernando Borretti because he's the only person who can explain artificial intelligence. I read him because I'm interested in how he thinks: his choice of subjects, his hostility toward certain ideas, the connections he makes, even the things he gets wrong.
Writing gives us an imperfect glimpse into another mind. Vocabulary, assumptions, awkward transitions, and the distinctions an author bothers to make all reveal something. An essay carries traces of the person who assembled it.
That creates a problem for AI-written essays. If I want something produced by ChatGPT, I already have ChatGPT. I can ask it for an argument, demand objections, or argue with it directly. What does someone else's generated article offer me?
Their ideas are a legitimate answer. A fresh chat doesn't know what I believe about machine consciousness or which distinctions matter to me. But the reader also wants to encounter the person who had those ideas, and a language model can obscure them.
The model interprets a discussion, chooses a structure, invents transitions, and supplies vocabulary. The arguments might be mine while the apparent intellectual personality belongs to neither of us. A faithful account of someone's thinking and a collection of vague observations dressed in sophisticated prose can look remarkably similar.
AI makes the distance between the author's thoughts and the reader difficult to determine.
The problem with artificial coherence
Language models are good at making mediocre thinking sound impressive. Feed one vague observations and questionable assumptions, and it can produce an essay more coherent than the ideas deserve. Where a human writer might stumble over a missing connection, the model can simply supply one.
A badly expressed argument announces its weaknesses. A beautifully expressed argument can hide them. Someone can now publish reasoning they couldn't reproduce, explain, or defend.
Smoothness also disguises the author. A reader might admire the apparent command of language and structure, then mistake those qualities for evidence about the person behind the text. This needn't be deliberate deception; the author may have used AI precisely because they struggled to express themselves. But the person implied by the prose may still be misleading.
And when people with different histories and temperaments all use the same linguistic machinery, they can begin to sound like variations of the same articulate stranger. I can make my ideas easier to understand while making myself harder to encounter. That's a loss, even when the tool is what made publication possible.
Thinking without writing
In Human Routers of Machine Words, Borretti challenges the defense that the human supplies the ideas while AI merely supplies the writing. Writing, he argues, is itself a process of thinking.
He's right. An argument that seemed persuasive in our heads can become contradictory on the page. A conclusion that felt inevitable suddenly requires premises we cannot justify. Expression tests whether we understand what we think we understand.
But must that test happen through the manual composition of prose? Discussing ideas can expose the same weaknesses. An intuition becomes a proposition; an objection forces a distinction; sometimes the original idea turns out to be nonsense. People have developed arguments through dialogue for millennia. For me, it's the more natural method.
AI is an imperfect partner for that work. It may be too agreeable, rescue a weak idea, or supply arguments I haven't earned. Asking it to manufacture an opinion and asking it to interrogate one are different activities, but merely intending the latter doesn't guarantee resistance. I still have to examine the result and understand why I accept it.
That leaves two separate questions: Did I think seriously about these ideas? And does reading the essay reveal how I think? The first can be true while the second remains doubtful.
There's also a limit to this defense. The reader cannot inspect the discussion that produced an article. They have the finished text and my account of a process they didn't observe. An AI-written essay can be intellectually sincere without providing reliable evidence of that sincerity.
The book he hated enough to read
Borretti's review of Marcus Hutter's Job-Less Utopia illustrates the distinction. Hutter acknowledged using Claude to polish his English and said he reviewed and approved the result. Borretti found the book verbose, repetitive, and burdened with unnecessary conceptual machinery. He suspected Claude had done more than polishing and announced that he'd treat it as the author.
Then he examined the arguments: automation, taxation, universal basic income, political power. He identified interesting claims, contradictions, and conclusions that didn't follow from their premises. He did what I'd hoped my friend would do. He engaged with the ideas.
There's no contradiction in despising AI prose while finding its arguments worth examining. Borretti could identify a failure of reasoning without knowing which sentence originated with Hutter. What he couldn't easily reconstruct was the thinking that produced it: had Hutter made the mistake, had Claude introduced it, or had Hutter approved something he hadn't adequately questioned?
The book could be a poor vehicle for encountering Hutter's mind while still containing ideas worth discussing. Those are different judgments.
Who owes whom an explanation?
The internet is filling with material that costs little to produce and still demands human time to evaluate. When an author publishes something they don't understand, they transfer the work of finding its unsupported claims and contradictions to the reader. That's an unreasonable exchange.
Disclosure helps. A reader deserves to know whether they're encountering someone's natural expression, an edited version, or a language model's reconstruction of a discussion. AI can shape the appearance of the reasoning as well as the words. These aren't identical experiences.
But knowing how something was produced doesn't settle whether its argument works. If a claim is unsupported, challenge it. If the logic fails, expose the failure. If the author can't defend what they've published, the suspicion of artificial coherence was justified. I'd rather discover that through engagement than assume it from the tool involved.
I also have to accept that readers may simply want a more personal voice. We don't necessarily want writing to be perfect. Sometimes we want it to be theirs. I read Borretti partly for that reason; I can't expect readers of my reply to want anything different.
The person behind the words
For me, the choice has often been between communicating through an imperfect intermediary and saying nothing. A window made slightly less transparent is still more revealing than a wall. AI has allowed thoughts that stayed in my head to be read, challenged, misunderstood, and developed by somebody else.
That doesn't entitle me to anyone's attention, or make the loss of my natural expression irrelevant. It gives me a reason to publish and a responsibility to stand behind what appears under my name.
I wanted my friend to ask what I thought. Perhaps he was asking exactly that. Just in a different way.