The spine

Sometimes it is language.

That matters to me because English is not my first language.

I can have an argument before I have the English required to carry it properly. I can know where the argument begins, what bothers me about the obvious answer, which example contradicts it, and where I think the conclusion may be hiding. I can explain all of that and still choose the wrong preposition, flatten a sentence, miss an idiom or construct a paragraph in a way a native speaker would not.

If a machine improves the English, what exactly has it done?

It may replace words.

It may move sentences.

It may tell me that the paragraph I thought was clear is not clear at all.

It may give me five ways of saying something and I choose the one closest to what I intended.

Sometimes it may propose a formulation better than mine.

Sometimes I reject it because the English is elegant but the idea has moved slightly away from what I mean.

By the time the article is published, a detector might find plenty of machine-shaped language in it.... But that tells us remarkably little about where the article came from.

We have dealt with versions of this problem for centuries.

Dostoevsky did not write The Brothers Karamazov in English. Most English-speaking readers encounter him through someone else’s sentences. A translator stands between the originating mind and the reader, making thousands of decisions Dostoevsky himself never made in English: which word carries the right weight, where a rhythm should survive, how awkwardness should be preserved rather than polished away, what an expression means when there is no exact equivalent.

Nobody therefore concludes that the translator wrote Dostoevsky’s novel.

The reason seems obvious.

The novel existed before the English.

The characters existed.

The architecture existed.

The conflicts existed.

The questions existed.

Something had already been made that needed to cross a linguistic border.

The translator helped it cross.

Translation is not identical to what artificial intelligence does, of course. A model can intervene much more deeply. It can propose structure, produce counterarguments, find examples, summarize research, suggest an analogy, or tell the writer that the premise itself may be wrong. It can become less like a translator and more like an editor, researcher or interlocutor.

But the old example reveals something useful.

Authorship has never required every visible word to originate directly from the person whose name sits above the work.

There is something underneath the words.

I have begun to think of it as the spine.

The spine is not the vocabulary.

It is not the grammar.

It is not whether the writer typed every sentence unaided.

It is the structure of intellectual responsibility running through the piece.

Why this question?

Why these examples?

Why does this section follow that one?

What would make me change my mind?

Which objection actually threatens the argument?

What am I uncertain about?

Why did I keep this paragraph and reject another one?

What do I believe after writing the piece that I did not quite believe before?

A writer may receive considerable assistance and still be able to answer those questions because the work remains connected to a mind that understands why it exists.

That is the distinction an AI detector cannot see.

Imagine two articles.

The first begins with someone who has spent years working in agriculture. He understands soil, seasons, machinery, prices, weather and the strange distance between how farming is discussed in cities and how decisions are made on a farm. He has an argument he wants to make, but his written English is poor. He talks to an AI system in his own language. He explains what he has seen. The system helps organize it. It asks questions. He corrects it. It translates the final piece into fluent English.

Most of the English sentences may have been produced by the machine.

But what exactly would we be detecting if we labelled that article artificial?

Now take another person.

He has never farmed.

He has read very little about farming.

He notices that regenerative agriculture is being discussed online and asks a machine:

Write me a nuanced Substack essay about what urban environmentalists misunderstand about farming. Make it sound grounded in lived experience but do not invent specific biographical facts.

The model produces something intelligent.

The person publishes it.

Perhaps the article contains fewer obviously machine-like phrases than the first one.

Which author is more artificial?

The question becomes almost embarrassing once stated that way.

The language is not where the difference lives.

The spine is.

And yet even that formulation needs care, because AI can participate in building the spine too.

Suppose I begin with a real question but no settled position.

I tell the machine what I think.

It gives me an objection.

I disagree.

It produces evidence that makes my objection weaker.

I read further.

The original argument changes.

A second possibility appears.

I reject it.

Then something in the rejection produces a third idea that neither the first prompt nor my initial position contained.

Who thought that?

The easy answers begin to fail.

It would be strange to claim that the machine contributed nothing. It would be equally strange to say that because the machine participated, the resulting thought is no longer mine.

Human thinking has always happened in contact with other minds.

A teacher asks a question.

A friend disagrees.

A book supplies an analogy.

A reviewer points to a contradiction.

A conversation turns an idea somewhere its author had not expected.

We do not normally say that a thought stops belonging to us because someone helped us reach it.

Perhaps AI can occupy part of that old territory.

The important difference may not be whether another intelligence entered the process.

It may be whether someone remained responsible for what happened inside it.

That is why I resist the phrase AI-written when it is used as though it described one activity.

It can describe radically different ones.

A machine can correct punctuation.

It can translate.

It can reorganize notes.

It can question an argument.

It can produce research leads.

It can draft a paragraph from a structure already decided.

It can propose the structure.

It can invent the thesis.

It can invent the personal hesitation around the thesis.

It can write the entire thing while the person whose name appears above it knows almost nothing about the subject.

All of that can be called AI-assisted writing.

The category is nearly useless.

What matters is not how much assistance occurred.

It is what was being assisted.

Was there knowledge that needed expression?

A question that needed testing?

An argument that needed structure?

A mind trying to cross a language barrier?

Or was the thing being assisted simply the desire to have something published?

That is why the spine matters.

But it also creates the problem that will follow us through the rest of this essay.

The spine cannot be counted.

It has no percentage.

There is no meaningful formula in which sixty-three percent human words makes something authentic and sixty-two percent does not.

Dostoevsky could contribute zero English words to an English translation and remain unmistakably the author of the novel.

Conversely, someone could manually rewrite every sentence of a machine-generated argument and contribute one hundred percent of the final keystrokes while contributing remarkably little to the thought.

Typing is not authorship.

Neither is fluency.

And perhaps, increasingly, neither is prose.

This matters far beyond writers who use English as a second language.

A scientist may understand a field profoundly and write badly.

A mechanic may know more about engines than someone with three degrees and have no idea how to construct an essay.

A nurse may have twenty years of observations about what happens around dying patients and no confidence in literary prose.

A migrant may think in one language and need to publish in another.

For all of them, artificial intelligence could do something genuinely liberating.

It could allow knowledge that already exists inside people to travel farther than their command of written language.

For centuries, translation allowed a mind to travel farther than its language.

AI may extend that privilege to people who never had translators.

That is not the destruction of writing.

It may be one of the largest expansions of who gets to participate in it.

Which is why any attempt to protect writing from AI by detecting machine-shaped sentences risks protecting exactly the wrong thing.

The polished native speaker passes.

The foreign writer who needed help is flagged.

The person who knows the subject but cannot structure paragraphs becomes suspicious.

The person who knows nothing learns to prompt the machine to sound less machine-like.

We would have built a gate against artificial intelligence and discovered that the people standing outside it were humans.

The more interesting question is therefore not whether AI touched the prose.

It is whether there is an accountable mind behind it.

Can the author explain the argument after the article is closed?

Can they tell you why an example matters?

Can they respond when somebody finds a contradiction?

Can they say, I don’t know?

Can they recognize when the machine has produced something elegant but wrong?

Can they reject a paragraph that sounds better than what they wrote because it no longer says what they mean?

Perhaps most importantly, can they stand behind the work when another person pushes back?

None of these are perfect tests.

A skilled performer can simulate them.

An AI can help answer the comments too.

There will be no clean border.

But they point toward something that word-level detection misses.

The article is not the whole object.

There is a relationship between the article and the person whose name appears above it.

That relationship is what authorship may increasingly have to mean.

The machine can help a thought cross a border.

It can help a thought discover that it was wrong.

It can even participate in making a new thought possible.

But there is another use entirely.

It can build the bridge when there was nothing waiting on the other side.

From the finished article alone, those two processes may look almost identical.

And that is where the problem stops belonging to the writer.

It becomes the platform’s problem.

Karim Spring