The detector looking at itself

The obvious response is to build a detector. If artificial intelligence can produce the article, perhaps another artificial intelligence can examine it and tell us what happened. The idea is reassuring because it restores the old categories. Human. Machine. Authentic. Generated. Put the text into a system, receive a probability, attach a label, and let the reader decide.

The problem is that the thing we are trying to detect is not entirely in the text.

Imagine the Thursday writer again. He asks a model to generate the subject, the thesis, the examples, the counterargument and the conclusion. The article arrives polished enough to publish. Before doing so, he asks the same model another question: Does this sound AI-generated? Perhaps it says yes. So he asks: Rewrite it to sound more human. Vary the rhythm. Remove generic transitions. Add some hesitation. Make the argument feel as though it developed while I was writing it. Avoid the expressions AI detectors tend to notice.

The model does that too.

Now the article goes to a detector. Another model examines the prose and tries to decide whether a model produced it. There is something almost comic about the arrangement. Artificial intelligence writes.... Artificial intelligence disguises the writing. Artificial intelligence inspects the disguise. Artificial intelligence assigns a probability to whether artificial intelligence was involved.

And all of them are looking at the same place: the sentences.

Where Intent Lives

But the question we actually care about occurred somewhere else—before the first sentence. Did the writer arrive with a thought and use the machine to help carry it? Or did the writer arrive with the need for a publication and ask the machine to supply the thought?

Those two histories can end in almost indistinguishable prose. No detector looking only at the final artifact can reliably reconstruct an intention that left no necessary mark on it. This is not simply a technical limitation that better models will eventually solve. It is partly an epistemological one.

There is no molecule of artificiality hidden inside a paragraph. A machine-generated sentence is not chemically different from a human-generated sentence. It is language. The words come from the same language. The syntax comes from the same language. The metaphors were learned from human writing. The rhythms were learned from human writing. The model became capable of sounding human precisely because humans supplied the examples from which it learned what human writing looks like.

A detector therefore searches for tendencies: predictability, regularity, certain distributions of words, patterns of phrasing, features more common in one set of texts than another. That may be useful in some contexts. It is not the same as discovering authorship.

• A human being can write mechanically.

• A machine can write irregularly.

• A non-native speaker can produce unusual constructions.

• A careful editor can smooth them away.

• A model can deliberately insert variation.

• A human can ask a model to make a paragraph less elegant.

The more models learn about the signals detectors use, the easier those signals become to alter. We should expect the problem to become harder, not easier.

The Eighty-Six Percent Dilemma

But suppose, for the sake of argument, that the technical problem were solved. Imagine a detector of extraordinary accuracy. It tells us that eighty-six percent of the visible prose in an article was produced by a language model.

What exactly have we learned?

Perhaps the author spent ten years studying the subject, dictated the argument in Arabic or Spanish, supplied the examples, challenged every conclusion and asked the model to produce fluent English. Eighty-six percent machine-produced prose.

Now imagine another article. The writer asks a model to generate the central thesis and structure. The machine provides the decisive analogy and the strongest objection. The human then rewrites every sentence manually. Zero percent machine-generated final prose.

Which article contains more human authorship?

The percentage has answered the least interesting question. We are counting words when what matters is responsibility. This is why translation remains such a useful disturbance to the categories. An English edition of Dostoevsky may contain no English sentence Dostoevsky ever wrote. A detector designed to identify who produced the visible words would correctly conclude: not Dostoevsky. And yet the conclusion would be useless. We do not ask whether Dostoevsky generated the English syntax. We ask what travelled through it.

The novel had a before. The translator received something. There was a world, a structure, a set of characters, tensions, questions and decisions already waiting to cross. Artificial intelligence complicates this because it can do much more than translate. But the logical problem remains. A detector can perhaps identify the hand that shaped the visible language. It cannot necessarily identify the origin of the intellectual object carried by that language.

Manufacturing Fingerprints

And AI creates an additional difficulty. It can manufacture not only the argument but the signs that an argument was discovered. This may matter more than style.

A serious essay often contains traces of movement:

• At first I thought...

• That answer turned out to be too simple.

• There is an objection here I cannot dismiss.

• I am less certain now.

• Perhaps the question is wrong.

Those sentences feel valuable because they appear to reveal the writer moving through uncertainty. They are fingerprints of thinking. Or they were.

A model has read enough essays to know that good thinking often leaves these fingerprints. So they can be requested: Make the argument evolve. Include a point where the narrator discovers that his original assumption was wrong. Do not make the conclusion obvious from the beginning. Add an objection strong enough to change the thesis. The model can now produce the literary shape of intellectual discovery without any corresponding discovery having occurred in the person who publishes it.

The old visual selfie had an equivalent technique. At first, an unplanned photograph looked unplanned because it was. Then we learned the visual grammar of spontaneity: Walk past again. Do not look at the camera. Laugh as though nobody is photographing you. Put the coffee there. Move slightly closer to the window. The final image contains all the evidence of a casual moment except the casual moment itself. Nothing in the photograph needs to be fake. The restaurant exists. The coffee exists. The person was really there. What was constructed was the relationship between the person and the moment.

AI can do something similar to writing. The books exist. The facts may be accurate. The argument may even be interesting. What can be constructed is the relationship between the named author and the intellectual journey the article appears to record.

The Gray Territory

That is why the question Was this written by AI? begins to feel increasingly primitive. It assumes there are two boxes: Human. Machine. But the important space is becoming the territory between them.

• A person can originate the thought and outsource much of the language.

• A person can originate the question and allow the machine to transform the answer.

• A person can argue with a model until neither the initial human position nor the initial machine response survives.

• A person can contribute almost nothing except the desire to have an article.

All four may produce polished text. A binary detector flattens relationships that are completely different.

Perhaps we will try disclosure instead. A small label: AI-assisted. That sounds better—until we ask what it means.

Spellcheck is AI-assisted. Translation can be AI-assisted. Research can be AI-assisted. Structural editing can be AI-assisted. The central argument can be AI-assisted. The entire article can be generated from one sentence. One label now covers everything from punctuation to intellectual substitution. It tells the reader almost nothing unless we describe the process underneath it.

And even then there is another problem. Honest writers will disclose. Dishonest ones may not. The writer using AI because English is not his first language may carefully explain the process and place a visible label on the work. The person manufacturing articles industrially may simply press Publish. A disclosure system can therefore become one more paradoxical mechanism in which the people most concerned with transparency make themselves look more artificial than the people who are not concerned at all.

The Limits of Platform Inspection

This is where the platform reaches the limit of inspection. It can inspect files. It can inspect patterns. It can inspect publication velocity. It can perhaps know when its own tools were used. What it cannot directly inspect is the thing we keep returning to: why the article exists.

Was there curiosity upstream? Knowledge? Experience? A question that would not leave? An argument that changed? Or merely a vacancy in the content calendar?

Intent has no reliable watermark. The machine can inspect the sentence. It cannot inspect the reason the sentence exists.

That does not mean platforms are helpless. It means they may be trying to solve the wrong problem if they make authorship primarily a forensic question. Perhaps the goal should not be to certify every article as sufficiently human. Perhaps the goal should be to make it harder for industrialized intellectual performance to overwhelm the conditions in which genuine writers and readers can find one another.

That is a different task. It does not ask Substack to look inside the writer’s head. It asks Substack to decide what kind of ecosystem it wants to build around the writing.

And that may be the point at which the problem becomes uncomfortable for Substack itself. Because the platform was built around removing gates. AI may force it to decide what happens when the gate remains open and the cost of producing what passes through it approaches zero.

Karim Spring