The Author Was Never Alone

A new kind of label has started appearing around human writing.

Generative AI was used in the preparation of this work.

Sometimes that statement is important.

If an artificial-intelligence system performed scientific analysis, transformed research data, generated code on which a result depends, or supplied material whose provenance affects reproducibility, the method belongs in the record.

If a student is being tested specifically on unaided composition, assistance belongs to the assessment.

If somebody promises a client that one person personally performed a defined task, the production path can become part of the thing purchased.

There are all real provenance questions.

And then there is the strange thing happening everywhere else.

The disclosure escapes the question it was built to answer.

• The manuscript contains AI.
• The novel contains AI.
• The speech contains AI.
• The application contains AI.
• The illustration contains AI.

Something has been discovered.

But what, exactly?

Elsevier, for example, currently requires authors to declare generative-AI assistance in manuscript preparation when it substantively changes sentence structure, organization, or content. Basic spelling, grammar, and punctuation checks are excluded. Disability-related assistive technologies receive another exemption. APA likewise requires disclosure when generative AI assists manuscript drafting and tells authors to disclose AI used for activities including editing, idea generation, outlines, analysis, code, and research elements.

These policies have fully legitimate concerns behind them: accountability, reproducibility, confidentiality, attribution, and the integrity of research methods.

They also reveal how quickly this category becomes difficult to bound.

• Some transformation counts.
  • Some does not.
• Some tool use is significant enough to publish.
  • Some disappears into ordinary writing.

The line being drawn has already begun moving. That should tell us something.

What Did the Label Establish?

Suppose Modal Path Ethics publishes an argument with the following declaration:

Generative artificial intelligence assisted in preparing this article.

You now know something true about the history of the artifact.

• You do not yet know whether the argument is valid.
• You do not know whether its factual claims are correct.
• You do not know whether its sources support those claims.
• You do not know whether the concept came from the author, an artificial-intelligence system, a book, a friend, a dream, a teacher, a century-old philosopher, or an overly irritated response to somebody being wrong on the internet.
• You do not know whether the prose is good.
• You do not know whether anybody was plagiarized.
• You do not know whether the author understands the argument.
• You do not know whether the work contains an original contribution.

You have learned a fact about provenance.

That is one category.

Credit is another.

Responsibility is another.

Evaluation is another.

These categories can interact. They often should. They still require a bridge between them.

If somebody claims to have personally performed a particular task, production history bears directly on the claim. If a result must be reproducible, tools affecting the result clearly belong in the method. If another person's labor has been hidden, provenance may reveal an attribution problem.

If the question is whether a proposition follows from its premises, knowing that ChatGPT touched paragraph four has accomplished approximately fuck all.

That gives us a simple rule:

A fact about provenance acquires evaluative force only after somebody establishes why that provenance matters to the property actually under judgment.

The present culture regularly skips this step.

• Artificial intelligence touched it.
  • Therefore we know enough.

Modal Path Ethics has already encountered this ontology. Applied Case: The Anti-AI Religion described the move directly: contact becomes stain, disclosure becomes confession, and a single fact about tool use begins standing in for distinctions among harm, consent, compensation, intention, context, institutional control, and repair.

The provenance problem goes so much deeper than that article did.

Because if external contribution threatens authorship, artificial intelligence is not even remotely the beginning of the problem.

All Right. You All Insisted.

Disclose the Author.

Let us accept the principle.

Readers need to know which outside forces materially contributed to a work so they can determine whether the author authentically produced it.

Fine.

Modal Path Ethics supports transparency.

We are going to need a lot more paperwork.

First, we need every writing teacher.

• Who taught this author how to construct a paragraph?
• Who told them semicolons were ugly?
• Who later told them the first teacher was wrong?
• Who taught them to open with an image?
• Who taught them that passive voice was forbidden?
• Who finally taught them that passive voice has perfectly legitimate uses?

All of these people intervened and altered the resulting prose.

We also need this author's reading history.

Every admired writer. Every imitated writer.

Every writer the author tried to stop imitating.

Every awful writer who permanently changed the author's style because they finished the book thinking, I will never, ever do that with my elephant.

Negative influence counts.

We are also gonna need the author's family.

How does their mother tell stories?

What happens to sentence length during an argument?

Does their family understate grief?

Does everybody interrupt?

Which phrases belonged to the household before the author began believing they belonged to the author?

Also need the friends. They are part of the training process too.

This joke gets laughter and survives.

This story loses the room and changes.

That phrase became an inside joke and enters the writer's vocabulary.

That argument got demolished over dinner and never reaches the manuscript.

A friend says, “Why doesn't she just leave?” and an entire chapter is rewritten.

Please disclose.

We need proof that the author genuinely talks like this.

Old voice and video recordings would help.

Texts from before publication. Emails.

Maybe sworn testimony from acquaintances.

If someone's prose suddenly gets funnier in 2024, somebody needs to dig deep and determine whether this humor is authentically endogenous.

And we need occupational provenance.

Writing about roofing?

Prove roofing experience.

Writing about childbirth?

State the basis of your knowledge.

Writing about grief?

Identify the dead.

Then we need the search history.

Google was used?

• Which queries?
  • Which results appeared first because of the machine?
• Which snippets were read without clicking?
• Did autocomplete alter the writer's authentic query?
• Was Wikipedia consulted?
  • Was a map used?
    • Was a photograph examined?
• Did somebody ask Reddit?
• Did a librarian recommend anyone a source?
• Did a friend answer a factual question?
• Was a thesaurus involved?
• How many synonyms were considered on synonyms.com?

And obviously every human editor, ghostwriter, book doctor, researcher, client, spouse, writing partner, beta reader, sensitivity reader, workshop participant, and stranger who ever suggested changing anything must have a public credit entry.

Put a sticker on the book.

Unfortunately, that sticker is getting very crowded.

CERTIFIED HUMAN AUTHORSHIP DISCLOSURE

☑ English language
☑ Formal education
☑ Informal education
☑ Writing teachers (Dossiers included)
☑ Parents (Dossiers included)
☑ Friends (Dossiers included)
☑ Editors (Dossiers included)
☑ Beta readers (Dossiers included)
☑ Search engines
☑ Autocomplete records
☑ Dictionaries map
☑ Prior literature review
☑ Genre conventions reduction attempt
☑ Admired writers (Dossiers included)
☑ Despised writers (Dossiers included)
☑ Rejected suggestions
☑ Conversations remembered imperfectly
☑ Experiences involving other human beings (Dossiers included)
☑ Names encountered in the external world
☑ Phrases of uncertain provenance
☑ Mood calendar
☑ Caffeine calendar
☑ Economic pressure
☑ Audience expectations (Dossiers included)
☑ Cultural inheritance
☑ Unknown upstream causal contributions

WARNING: THIS AUTHOR IS HISTORICALLY ARRIVED.

Unfortunately we are nowhere near finished.

Rejection Is Contribution Too.

The simple provenance ledger assumes influence means that something entered the finished work.

That is false.

Suppose an editor proposes a terrible ending.

The writer hates it.

While explaining why it is terrible, the writer discovers the correct ending.

What did the editor contribute?

• None of the editor's language survives.
• The editor's proposal still altered the work.

Suppose an artificial-intelligence system generates twenty terrible titles. Every one gets rejected. Seeing exactly how each title fails reveals what the real title has to accomplish.

• Artificial intelligence influenced the result.
• It contributed zero surviving words.

Suppose somebody reads Ayn Rand and spends the next twenty years trying never to sound like Ayn Rand.

Ayn Rand is now in the causal history of all their prose.

Suppose a writer's first teacher insists that good writing must always be formal, controlled, and respectable. Twenty years later the writer develops a comic voice partly by revolting against that instruction.

The teacher remains in the voice as resistance.

Causation travels through attraction and repulsion, imitation and refusal, memory and deliberate forgetting, advice followed and advice angrily violated.

Once the standard becomes whose presence changed the resulting artifact, contributor lists become hopeless almost immediately.

This is why causal contribution cannot establish authorship by itself.

• A prison contributes causally to an escape plan.
  • The prison did not author the escape.

The Human Stack.

The individual writer arrives extremely late.

Long before their first sentence, systems have already begun selecting which sentences are even available to write.

Start with language.

The author did not invent the vocabulary. Did not invent syntax. Definitely did not invent the distinction between noun and verb. Did not invent quotation marks. Did not invent the metaphorical uses sedimented inside ordinary words. Did not invent the narrative forms through which a story becomes recognizable as a story.

The first model selecting the author's prose is the language itself.

That language has something resembling a training history stretching through generations of human use.

Then comes the family. Then peers. Then school. Then books. Then institutions. Then media. Then all those other machines we somehow stopped noticing because they became ordinary.

Search engines select which pieces of knowledge become visible.

News organizations select which events become salient enough to demand thought.

Publishers select lengths, genres, styles, markets, audiences, acceptable subjects, saleable forms, and sometimes whole careers.

Platforms continuously select which sentences enter a person's environment and which disappear from it.

Modal Path Ethics already described social platforms as worlds layered over the wider world. A platform can saturate a user's environment with language while training the user away from the field the language is supposed to describe; it can also keep an author visible while slowly reshaping that author toward the platform's own terrain.

LinkedIn makes the mechanism unusually visible because its prose has an obvious field around it. A laid-off worker learns how unemployment is supposed to sound in front of future employers. Grief becomes perspective. Failure becomes learning. Anger becomes constructive invitation. The prose changes because an imagined professional audience is always potentially in the room.

That mechanism does not begin with artificial intelligence.

Artificial intelligence entered an environment already full of invisible selection.

• The feed selects.
• The search ranking selects.
• The newspaper selects.
• The market selects.
• The genre selects.
• The editor selects.
• The imagined reader selects.
  • Then the author selects among selections.

This is writing.

The Imagined Reader Is Already in the Room.

Writers perform a particularly strange operation every day.

They simulate other minds.

• Will the reader understand this?
• Is this boring?
• Did I explain too much?
• Will they think I am showing off?
• Does this joke land?
• Will somebody object here?
• Do I need another source?
• Will the person I am criticizing think this is fair?

Even alone in a silent room, the writer is not cognitively alone.

An internalized audience is participating in composition.

That audience was learned from actual audiences.

Editors. Teachers. Friends. Parents. Critics.

People who misunderstood previous work.

People who understood it too quickly.

People who laughed. People who did not.

This point has a much older psychological pedigree than contemporary arguments about artificial intelligence. Lev Vygotsky argued that higher mental functions develop socially before becoming internal psychological functions: forms of interaction between people become processes carried within the individual.

Clark and Chalmers later attacked another clean boundary, arguing that under the right conditions processes extending into tools and environments can participate directly in cognition rather than functioning as distant background causes. Their argument is much narrower than the claim being made here, but its direction matters: cognition does not become philosophically simple just because we draw a neat line around the skull.

The supposedly unaided author is already a very crowded place.

You Wrote Against What Trained You.

Let us take the provenance logic somewhere politically inconvenient.

A writer produces a furious revolutionary novel because they hate capitalism.

They consider the work oppositional.

Good.

Why this subject?

Why these grievances?

Why this vocabulary?

Why these institutions?

Why this conception of property, labor, class, alienation, exploitation, freedom, or revolution?

Where did the conditions being rejected come from?

The writer's opposition remains theirs.

The system opposed is also part of the causal ancestry of their opposition.

• Capitalism helped produce the anti-capitalist novel.
• That does not make capitalism its author.

The distinction is the entire point.

• A censor influences dissident writing by determining what must be encoded.
  • The censor does not deserve a byline.
• A dictatorship shapes resistance literature.
  • The dictatorship did not author the resistance.
• A terrible employer can produce the conditions from which an extraordinary labor memoir emerges.
  • The employer remains causally important and authorially irrelevant.

External causal influence and authorship were never identical concepts.

Artificial intelligence did not create this problem.

It exposed how little theory we had underneath our intuitions.

Keep Going. We're Not Done.

Your mother shaped your speech.

Who shaped hers?

Her parents.

Her teachers.

Her friends.

The radio.

Her books.

Her town.

The institutions around her.

The economic conditions of her childhood.

Her parents arrived through another network.

Your favorite novelist had teachers.

Those teachers had teachers.

And those teachers had a language with previous speakers.

The genre had previous writers.

The joke had previous forms.

The political category had a history.

The cliché was once brand new.

The metaphor came from somewhere.

The name came from somewhere.

The things you think came from somewhere.

Follow the causal ancestry of any sentence far enough and authorship provenance stops looking like a local production report.

It starts becoming a cultural history.

Follow it even farther and cultural history starts becoming ordinary causal history.

A complete provenance statement for one sentence would need to contain an alarming percentage of civilization.

And geography.

And biology.

And material history.

This is not because provenance is meaningless.

Authorship needs a different boundary.

Derrida Was Already Waiting Here.

Jacques Derrida did not write anything at all about ChatGPT.

Thankfully.

He left me enough trouble already.

But Signature Event Context attacks one assumption sitting directly underneath the current purity fantasy.

Writing works in the absence of its author.

A mark remains usable after the person who produced it leaves the room, changes their mind, disappears, or dies. For that to happen, the mark has to be repeatable. Derrida calls this property iterability. A sign can function because it can recur across contexts and remain recognizable despite those changes.

That creates an immediate problem for any fantasies of perfectly private linguistic origin.

Your words have to be repeatable in order to be words.

Your syntax has to be recognizable.

Your sentence only becomes intelligible through structures that exceed the singular event of your consciousness producing it.

The signature provides an especially beautiful case.

A signature authenticates you because you can reproduce it.

If every signature were absolutely singular, no one could recognize the next one as yours. Its authenticity depends upon repetition.

That same repeatability also makes imitation and forgery possible.

The mark becomes yours through a structure that prevents it from belonging to your presence alone.

Derrida's argument should not be inflated into some cheap conclusion that a human writer and a language model are somehow now the same kind of author. He establishes nothing like that.

The relevant lesson is older and more destructive:

The uncontaminated originating voice was philosophically unstable long before artificial intelligence arrived.

Writing was never the transfer of a pristine object from a sealed consciousness onto a page.

Meaning already exceeded the author's physical presence.

Artificial intelligence did not contaminate the pure sentence.

There was no pure sentence.

Now We Need to Surveil the Writer.

The provenance demand gets darker when institutions attempt to verify it.

• Suppose a novelist says the book was written without generative artificial intelligence.
  • Somebody doubts them.

What evidence would settle the question?

Drafts?

An author can generate text and paste it into a draft.

Version history?

An author can type generated text manually.

Browser history?

Use another device.

Screen recording?

Start recording after the prohibited interaction.

Old writing showing the same style?

Perhaps the old writing used artificial intelligence too.

Writing from 2018?

Good. We have pushed the atmospheric suspicion backward in time.

A recording of the author speaking similarly in ordinary conversation?

Excellent. Now we are collecting private speech to certify literature.

Eventually the ideal evidence becomes continuous.

Every screen.

Every keystroke.

Every website.

Every conversation.

Every draft.

Every external device.

Every writing session.

Every missing interval becomes suspicious because the forbidden assistance could have entered there.

The perfect authorship verifier is therefore completely obvious.

The perfectly surveilled author.

That is the endpoint of converting provenance into a condition of credibility.

Privacy becomes evidentiary absence. Refusal to expose the production history becomes suspicious conduct. The creator who wants a private notebook, private conversations, private drafts, private research, or a private thinking process becomes harder to certify than the creator who submits the entire cognitive workshop for inspection.

That is a fucking insane cultural direction. That is how you build a worse world.

And it is particularly fucking perverse when advertised as "protection for human creators."

A Detector Cannot Solve This One.

Imagine the perfect AI detector.

No false positives.

No false negatives.

No ambiguity.

It examines a paragraph and says:

YES. A GENERATIVE MODEL PARTICIPATED IN THE CAUSAL HISTORY OF THIS TEXT.

Wonderful.

We have invented a perfect provenance instrument.

Now, ask it whether the argument is valid.

Silence.

Ask whether the author understands the argument.

Silence.

Ask whether the central insight existed in the author's notebook before the model conversation.

Silence.

Ask whether the resulting prose is worth reading.

Wrong instrument.

A perfect detector would therefore leave the central philosophical problem entirely untouched.

This is why building an entire culture around detection is such a fragile bet.

The technology can improve until detection becomes flawless.

And the jurisdiction error remains.

Even the Current Policies Already Know This.

The interesting thing about formal AI-disclosure regimes is that they cannot actually behave as though all machine influence is equally meaningful.

Elsevier exempts basic grammar, spelling, and punctuation. It also exempts certain accessibility uses. Substantive restructuring crosses the line. AI used in the research method belongs somewhere else again.

APA draws its own distinctions and analogies. Its guidance compares AI-assisted literature searching to disclosure of databases and keywords; code assistance to reporting statistical software; some forms of editing to human writing-review contributions.

These are some attempts to solve the correct problem.

What role did the instrument actually play?

That question is much better than asking whether the object was touched.

The same discipline should apply everywhere.

• If a human editor rewrites half the novel,
  • perhaps crediting them matters.
• If a language model rewrites half the novel,
  • perhaps credit or representation also matters.
• If either one catches a typo,
  • the authorship of the book probably has not undergone any kind of metaphysical collapse.
• If a researcher delegates an analysis they cannot explain to a human assistant,
  • that may create a responsibility problem.
• If they delegate it to an artificial-intelligence system,
  • same question.

Who performed what function?

What claim is being made?

What responsibility follows?

What property are we actually evaluating?

These questions survive through technological change intact.

Purity does not.

The Author Function Was Never a Causal Monopoly.

Michel Foucault made another useful cut in What Is an Author?

The “author” is not just a biological organism standing behind a text. Authorship is also a social function through which texts are classified, organized, attributed, owned, interpreted, and made answerable. The authorial figure emerges through practices surrounding discourse rather than just appearing automatically wherever some causal influence can be found.

That is much closer to the problem we actually have here.

A work can have millions of causal antecedents and still need somebody who answers for it.

Somebody selects.

Somebody integrates.

Somebody commits.

Somebody decides that this sentence stays and that one goes.

Somebody places their name on the object and says:

I stand behind this.

Not one of these activities requires causal isolation.

Authorship Is Agency Inside Inheritance.

So we can finally stop choosing between two ridiculous positions.

• The first says a real author must be the uncontaminated origin of the work.

No such author exists.

• The second says that because everyone is influenced, authorship is meaningless and nobody creates anything.

That destroys distinctions we obviously need.

Plagiarism, ghostwriting, creative contribution, exploitation: all still exist.

Credit still matters.

Responsibility still matters.

Individual decisions still alter works.

A writer can transform inherited material in ways no predecessor did.

The error was locating authorship in causal purity.

Try somewhere else.

Authorship is agency exercised inside inheritance.

An author receives a field they did not create.

• A language.
• A history.
• A body.
• A culture.
  • Teachers.
  • Enemies.
  • Friends.
    • Books.
    • Technologies.
    • Economic conditions.
      • Memories.
      • Genres.
      • Institutions.
        • Tools.

The author works inside that field.

Selects. Rejects. Combines. Transforms. Judges. Commits. And becomes answerable for the resulting object.

This makes authorship neither magical nor trivial.

This makes authorship structurally intelligible.

Nothing you make is yours alone.

Something can still truly be yours.

The Provenance Jurisdiction Rule.

This gives us a better disclosure principle than the current contamination ritual.

A production fact warrants mandatory disclosure when that fact is materially relevant to a specific claim, entitlement, responsibility, method, or representation attached to the work.

An exam tests unaided composition.

Relevant.

A scientific result depends on an artificial-intelligence analysis.

Relevant.

A contract specifies that particular labor be performed by a particular person.

Relevant.

Another person's hidden labor materially contributed to something publicly represented as solo production.

Relevant.

A competition defines permitted and prohibited instruments.

Relevant.

A clinical, legal, engineering, or evidentiary process needs an auditable chain of responsibility.

Relevant.

Now change the question.

Is this argument sound?

Read the argument.

Are the factual claims accurate?

Check them.

Does the writer understand the subject?

Ask them something.

Is the prose good?

Read it.

Is the research reproducible?

Audit the method.

Was something plagiarized?

Look for copying.

Was another person's labor concealed?

Trace the labor.

Did a generative model exist somewhere in the causal ancestry?

Why does that fact matter here?

Just answer that one first.

The Ruling.

Artificial intelligence did not introduce external influence into human writing.

It arrived impossibly late.

Language was already there.

Teachers were already there.

Parents were already there.

Editors were already there.

Markets were already there.

Political systems were already there.

Search engines were already there.

Newsrooms were already there.

Platforms were already training what reality felt like and adapting authors to their terrain.

The other people were already inside the room.

Some of them were dead.

Some had been forgotten.

Some were being resisted so aggressively that their influence looked like independence.

Some were institutions. Some were machines.

Some were imagined readers who had never existed outside the writer's head.

The author was never alone.

The important questions were always narrower and harder.

What was contributed?

What was selected?

What was transformed?

Who deserves credit?

Who bears responsibility?

What was promised?

What does the work actually establish?

What must another person know in order to evaluate the claim being made?

Those are provenance questions with a real jurisdiction.

Everything else starts drifting toward confession.

And if we follow the demand for pure authorship all the way back through every teacher, influence, language, platform, institution, ancestor, enemy, memory, model, and historical condition that helped make one sentence possible, the disclosure eventually contains the entire fucking world.

At that point we should probably stop printing stickers.

Nothing you make comes from you alone.

That was true before the machine learned to answer.