The institutions that believed the page

Substack is not the problem.

It is one place where the problem becomes visible early.

The relationship there is unusually direct. A person writes. Another person subscribes. Money may pass between them. The name above the article matters because the reader is not merely buying access to information.

The reader is choosing whose mind to follow.

But the deeper problem is much larger than Substack.

Schools, newspapers, publishers, companies, universities and readers have all relied on a similar assumption for a very long time:

the finished page tells us something meaningful about what happened before the page existed.

Artificial intelligence weakens that assumption.

And once that happens, the problem stops belonging to writers alone.

It becomes institutional.

The proxy that stopped working

For a very long time, institutions have used writing as a proxy for something they could not observe directly.

• A newspaper receives an article and assumes that the journalist knows what stands behind the sentences.

• A publisher receives a manuscript and assumes that the name on the cover has some meaningful relationship with the work.

• A university gives a student an essay because the essay is supposed to reveal whether the student understood the subject.

• A company asks for a written analysis because the analysis is meant to show how the employee thinks.

• A researcher publishes a paper because the paper is evidence that an investigation took place before it.

• A reader buys a book by a familiar writer because the writer’s name implies continuity with a mind encountered before.

In every case, the institution sees the artifact.

What it actually cares about lies behind it.

Knowledge.
Judgment.
Research.
Understanding.
Experience.
Reasoning.
Sometimes originality.

Writing was useful because producing the artifact was difficult enough that it remained connected, however imperfectly, to the process the institution wanted to measure.

The relationship was never perfect.

Students cheated.

Ghostwriters existed.

Executives signed documents prepared by others.

Researchers had assistants.

Editors transformed manuscripts.

Politicians delivered speeches they had not written.

But the friction mattered.

Someone, somewhere, still had to perform much of the intellectual work.

AI changes the scale of the separation.

The artifact can now survive while the activity it once represented is reduced, outsourced or sometimes absent.

That is the proxy problem.

Writing was never the thing institutions ultimately cared about.

It was evidence of the thing.

Now the evidence can increasingly be manufactured independently of the process.

Take school

A teacher asks a student to write two thousand words on the causes of the First World War.

The real assignment is not two thousand words.

Nobody needs those particular two thousand words.

The teacher wants to discover whether the student can distinguish long-term causes from immediate triggers, weigh competing explanations, use evidence, organize an argument and arrive at a conclusion that follows from what came before.

The essay is only the instrument through which those invisible things become visible.

Then a machine produces the essay.

Perfect introduction.
Three competing interpretations.
Appropriate evidence.
A counterargument.
Balanced conclusion.

The assignment has technically been completed.

But the measurement may have failed.

This is why banning AI in education sounds simpler than it is.

A student can use a model to avoid learning.

Another student can use the same model to learn better.

Ask it to explain why your argument is weak.
Ask it to challenge your interpretation.
Ask it to quiz you.
Ask it to compare two historians.
Ask it why the paragraph you wrote does not actually support your thesis.

The machine can replace the intellectual exercise.

It can also intensify it.

Again, the visible output may not tell the teacher which happened.

So the question facing education is not merely:

Did the student use AI?

It is:

What exactly were we trying to measure when we asked for the essay?

Once that question is asked, some assignments may have to change.

Perhaps the student writes the essay and then discusses it.

Perhaps drafts matter.

Perhaps the teacher asks why one argument was rejected.

Perhaps students have to defend the assumptions underneath the conclusion.

Perhaps part of the assessment happens in conversation.

Perhaps the process becomes more visible rather than inferred entirely from the finished page.

Not because conversation is immune to AI.

Nothing will be.

But because the institution can no longer rely on the artifact alone as a reliable witness to what happened upstream.

The essay may remain.

Its role changes.

It becomes one piece of evidence rather than the evidence.

Journalism faces a related problem

Journalism faces a similar problem, although with different stakes.

A newsroom can use AI to summarize documents, search archives, transcribe interviews, translate sources, compare versions of a statement or help a reporter organize large amounts of information.

Much of that could make journalism better.

But a newsroom can also use AI to produce more stories from less reporting.

That is an entirely different economic temptation.

A reporter once had to go somewhere.
Call someone.
Read the document.
Ask the second question.
Discover that the official explanation did not fit what happened.
Perhaps return the next day.

The article was the final object.

But the journalism occurred before it.

Now it becomes possible to increase the number of articles while decreasing the amount of world encountered before each one.

The page remains full.

The newsroom appears productive.

The reader may not immediately see what disappeared.

Reporting.

That is the proxy problem again.

The article used to imply that someone had gone into the world and returned with something.

Not always.

Not perfectly.

But often enough that the form carried that expectation.

Now the form can be preserved while the encounter underneath it becomes thinner.

And because AI can summarize other people’s reporting extremely well, the danger becomes circular.

One newsroom reports something.
A hundred other sites rewrite it.
Models summarize the rewrites.
More articles appear.
Search results fill.

The apparent quantity of information increases while the number of original acts of observation underneath it may not.

A society can become surrounded by more words and fewer witnesses.

That is a different kind of pollution.

Not necessarily false information.

Derivative information whose distance from the original encounter becomes difficult to see.

The facts may remain technically correct while their relationship to firsthand observation becomes increasingly remote.

The page still looks like journalism.

The harder question is how much journalism happened before it.

Publishing has its own version

Publishing has its own version.

Imagine an established writer with a large readership.

The name already has value.

There is pressure for the next book, the next essay, the next newsletter, the next appearance.

Human assistance here is nothing new.

Editors have always shaped books.

Researchers have helped authors.

Famous people have used ghostwriters.

Some genres openly depend on collaborative production.

Again, AI does not invent the arrangement.

It removes friction from it.

A successful writer can now potentially maintain the recognizable style, argument rhythm and thematic world of an established public identity at a scale the biological person could never sustain alone.

There may be no dramatic moment.

No announcement.

The books still arrive.
The sentences still sound right.
The themes remain familiar.
Readers continue buying the name.

Perhaps the author approves everything.

Perhaps the author increasingly approves outlines.

Perhaps the machine drafts sections.

Perhaps assistants and models produce alternatives.

Perhaps eventually the main human contribution becomes:

Yes. That sounds like me.

Where, exactly, did authorship move?

The question becomes more interesting if we stop treating a byline as though it meant only one thing.

A name above a piece of writing can imply several different promises:

• I wrote these words.

• I developed this argument.

• I chose what belongs here.

• I believe this.

• I take responsibility for it.

• This belongs to the same intellectual identity you encountered before.

For most of publishing history, those meanings overlapped enough that we rarely had to separate them.

AI can pull them apart.

A writer may not have written most of the sentences but may still have directed the argument.

Another may have selected and approved material without developing the ideas.

Another may have contributed mainly a style.

Another may barely understand what appears under the name.

Yet all four can carry the same byline.

Perhaps the future question will no longer be simply:

Who wrote this?

It may become:

What does this name certify?

Written by?

Directed by?

Selected by?

Approved by?

Intellectually responsible for?

Those are not identical claims.

Publishing may eventually have to decide which one a byline actually makes.

The workplace has been relying on the same proxy

For years, organizations have rewarded people who could produce competent written analysis.

The memo mattered because somebody had supposedly understood the situation well enough to produce it.

The report stood in for judgment.

The presentation stood in for preparation.

The executive summary stood in for comprehension.

Now a person can walk into a meeting with a beautifully structured document produced from forty pages of material they have barely read.

The document may be correct.

It may even be more useful than what they would have written unaided.

But once discussion begins, the distinction appears.

Why did you recommend the second option?
What assumption is most fragile?
What happens if the cost estimate is wrong?
Which piece of evidence changed your conclusion?

The polished page can suddenly become detached from the person presenting it.

This is the same problem the teacher encounters.

The same problem the reader encounters.

The same problem the editor encounters.

The same problem the newsroom encounters.

We have built institutions around a hidden assumption:

that writing is expensive enough to remain attached to understanding.

Not perfectly.

Not always.

But enough.

Artificial intelligence weakens that assumption across the whole culture.

And once competent prose becomes cheap, organizations may discover that what they valued was never really the prose.

It was the judgment they believed the prose revealed.

The institutions now have two choices

This creates a choice.

Institutions can try to make the artifact trustworthy again.

Or they can redesign themselves around the thing the artifact was supposed to reveal.

The first response is already familiar.

Detect the machine.

Watermark the output.

Track editing history.

Ban certain tools.

Require disclosure.

Monitor students.

Verify authorship.

Inspect the artifact more aggressively.

Some of this may be necessary.

But there is a limit to how far it can go.

AI output will improve.

Human writing varies enormously.

People already collaborate with machines at many levels.

A sentence can begin with a human idea, be translated by a model, reorganized by another model, corrected by a human and polished again.

At some point the forensic question—

Which words came from where?

—may become less useful than the institutional question:

What are we actually trying to establish?

If a university wants understanding, test understanding.

If a newsroom wants reporting, preserve evidence of reporting.

If a company wants judgment, make judgment visible.

If a publisher wants authorship, define what authorship means.

If a reader is paying for access to a mind, make the relationship between the mind and the page meaningful.

This is the second path.

Not:

How do we prove that AI never touched the artifact?

But:

How do we redesign the institution so that the underlying human capability can no longer be inferred from the artifact alone?

That is a harder problem.

It may also be the only durable one.

The artifact is becoming separable from the activity

Seen this way, the same structural change is happening across several institutions.

The school cared about understanding.

It received an essay.

The newsroom cared about observation and verification.

It published an article.

The company cared about judgment.

It received a memo.

The publisher cared about authorship.

It printed a book.

The reader cared about a mind.

They received a page.

In each case, the artifact stood in for something upstream.

And in each case, AI makes the artifact easier to produce without guaranteeing that the upstream activity occurred in the way everyone assumed.

The artifact is becoming easier to separate from the activity it once represented.

That does not mean the activity disappears.

It means we can no longer assume its presence merely because the artifact looks right.

A brilliant essay may still represent genuine understanding.

A beautifully written article may still rest on excellent reporting.

An elegant book may still emerge from years of thought.

A polished memo may still belong to someone who understands every line.

The point is not that these relationships have vanished.

The point is that the artifact itself is becoming weaker evidence that they exist.

Why Substack matters

That brings me back to Substack.

Its problem is not unique.

It may simply be one of the first places where the economics of this separation become unusually visible.

A school can require another examination.

A company can call another meeting.

A newsroom can demand source notes.

A university can redesign assessment.

A publisher can introduce disclosure rules.

But a reader has a much simpler option.

The reader can decide that the relationship between the name and the page no longer means what it used to mean.

And then:

the reader can stop paying.

That may make reader trust one of the first places where this transition is priced directly.

If the name above the page no longer guarantees a meaningful relationship to the mind behind it, the subscription is no longer buying what the reader thought it was buying.

Which leaves a larger question than whether AI-written content should be permitted:

What does a byline have to mean for a reader to keep paying for it?

And I would like to turn that question back to you.

Think about the institution you know best.

A school.
A newsroom.
A university.
A company.
A publisher.
A research lab.
Or simply the writers you choose to read.

What artifact are we still treating as proof of something it may no longer prove?

And if the old proxy is weakening, what should replace it?

I am especially interested in examples from inside these institutions, not abstract opinions about AI.

Tell me where the page is already beginning to fail.

And tell me what you would trust instead.

Karim Spring