Substack Says I Am 100% Artificial. That Will Be News to Me.

This morning, I finished an article about paperclips. Not really about paperclips.

It was about all the tiny operational workarounds a business learns to tolerate because each one seems easy enough.

“Just jiggle the handle.” “It’s in Google Drive somewhere.” “Ask Susan. She knows where we keep it.” “Put a paperclip under it.”

The article began with a real process inside my real company involving a real coffee program we have operated every month for more than five years.

The story was mine. The experience was mine. The framework was mine. The argument was mine. The examples came from my life and work. The final editorial decisions were mine.

Then I asked Substack’s new AI-text scanner to assess it.

The result?

AI: 100%
AI-assisted: 0%
Human: 0%

Apparently, I was not there.

Which is fascinating, because I distinctly remember the entire conversation.

Substack recently added an optional “Scan for AI text” function powered by Pangram. It can be run on posts and Notes published on or after July 21, 2026, through the Substack Reader. Substack describes the result as an estimate of whether text is human-written or AI-assisted.

Mine did not come back “AI-assisted.”

It did not come back “mixed.”

It did not detect even a faint pulse of humanity.

It classified the article as fully AI-assisted text and assigned:

100% AI.
0% AI-assisted.
0% human.

Well.

Someone should probably tell my nervous system that it did not participate.

It will be relieved. But I am actually pissed.

I write through conversation.

I talk through an idea. I argue with it. I follow it somewhere unexpected. I pull examples from business, law, coffee, public safety, military capability development, family enterprise, and the occasional migratory duck.

• I identify the point beneath the story.

• I decide what I believe.

• I test the language.

• I reject language that is not mine.

• I add things.

• I remove things.

• I rewrite transitions.

• I change examples.

• I put back sentences that an editor removed because I liked them better the first time.

• I read the piece aloud.

• I revise it again.

And yes, I use AI assistance.

Openly. Deliberately. As a conversational drafting and editorial tool.

That is not the same thing as entering: “Write me an article about business.”

Then making a sandwich while the machine invents my work.

The distinction matters.

My writing process now includes AI.

Before that, it included editors.

Before that, colleagues.

Before that, whiteboards, interviews, notebooks, recorded conversations, legal research, peer review, source material, subject-matter experts, and long arguments with myself at unreasonable hours.

Writers have always used tools and other minds to develop work.

The presence of assistance does not erase the author.

It does not erase the origin of the idea.

It does not erase intellectual property.

It does not erase lived experience.

It does not erase editorial control.

And it certainly does not establish that there was zero human input.

Pangram itself distinguishes among human-written, AI-assisted, and AI-generated text. Its publishing materials describe AI-assisted work as work in which a human author used AI tools during the writing process, while its API separately classifies text as human-written, AI-assisted, or AI-generated.

That is exactly why a result of 100% AI and 0% AI-assisted matters.

It does not describe my process.

Not approximately. Not loosely. Not at all.

I have written in this voice for more than thirteen years. Long before generative AI writing tools became publicly available.

I wrote and published through At The Ready Magazine starting in 2013. Look at those articles.

I wrote and published articles, editorials, business commentary, professional frameworks, podcast scripts, legal analysis, course material, and books. The first book was in 2021. Same voice, style, and yes; bullet points, snarky asides, and short, easy to digest ideas.

• The cadence did not suddenly arrive with AI.

• The short paragraphs did not arrive with ChatGPT.

• The repetition did not arrive with Artificial Intelligence.

• The sharp turn from ridiculous story to serious operating principle did not arrive with “AI Assistance”.

The parenthetical side comments, escalating examples, slightly feral metaphors, and occasional “ahem” are not evidence that a machine invented Dawn Kennedy.

They are evidence that Dawn Kennedy has been writing like Dawn Kennedy for a very long time.

I already have more than 75 Substack articles in the same established voice.

A longitudinal record exists. But the detector does not examine that history.

It examines the finished text.

It does not know when my voice developed, where the idea came from, which sentence began as something I said aloud.

And it does not know which argument reflects more than eleven years of decision-architecture work.

It does not know me.

Yet its result appears in the confident language of measurement:

“AI Generated -100%.”

That number looks far more certain than the system’s access to the evidence allows.

Pangram says its detector analyzes patterns in text to classify AI involvement. It also acknowledges the existence of false positives, which occur when human writing is classified as AI-generated.

The company reports very high accuracy and a very low false-positive rate for its system.

Fine.

But a low error rate does not comfort the person standing inside the error.

More importantly, text-pattern analysis and authorship analysis are not the same inquiry.

A detector may estimate that the linguistic construction of finished prose resembles text associated with generative AI.

It cannot determine, from the finished text alone:

• Who originated the idea?

• Who owns the framework?

• Who supplied the facts and lived experience?

• Who directed the argument?

• Who challenged the draft?

• Who selected the examples?

• Who made the final editorial decisions?

• Who has used the same voice for thirteen years?

• Who is legally and professionally accountable for the published work?

Those are authorship questions.

A percentage generated from prose patterns does not answer them.

I use AI assistance. There is no scandal there.

I am not pretending I compose every article alone in a candlelit room using a fountain pen and suffering.

Although I am an author, so suffering remains available.

The problem is that the result erased the distinction Substack and Pangram claim to measure.

It did not say:

This author appears to have used AI during drafting or editing.

Fair enough.

It said, functionally: The text is entirely AI, with no detectable human contribution.

That is a materially different characterization.

My concern is not embarrassment over using a modern tool. My concern is the reduction of authorship to a detector score that cannot see the process it is purporting to characterize.

Because readers may treat the result as proof.

Editors may treat it as proof.

Clients may treat it as proof.

Search engines, platforms, schools, employers, publishers, and professional communities may increasingly treat these scores as proof.

And “the model is usually accurate” will not restore the reputation of the individual it describes inaccurately.

Substack offers authors the opportunity to create a “How I make this” statement.

So here is mine:

Dawn develops her work through lived experience, original intellectual property, conversation, argument, revision, and final editorial judgment, with AI assistance. Her voice and cadence are documented across more than thirteen years of published writing, beginning with At The Ready Magazine, long before generative AI writing tools entered common use.

This statement appears here because, apparently, authors must now explain to an algorithm how authorship works.

I resent needing that statement.

Not because it is untrue.

Because the burden has somehow shifted.

A detector produces an estimate based on the text alone.

The author is then expected to explain why her documented career, ideas, voice, revision process, and editorial judgment should count as human participation.

That is backward. The score should carry the disclaimer.

Not the author.

I am not arguing that platforms should never identify undisclosed bulk-generated content.

I understand the concern.

Readers deserve to know whether a person is presenting machine-produced material as reporting, expertise, lived experience, or original analysis when it is none of those things.

But responsible transparency must include transparency about the detector.

A result should make clear that:

• It is an estimate.

• It analyzes the finished text, not the complete authorship process.

• It cannot determine ownership of ideas or intellectual property.

• It cannot establish the absence of human direction, revision, or editorial judgment.

• It may produce false positives or misclassify the degree of assistance.

• It should not be treated as definitive proof of deception.

• And there needs to be a meaningful way for an author to dispute or contextualize a result.

Otherwise, the tool does not create transparency.

It creates a new form of opacity dressed as a percentage.

I keep returning to that part.

Human: 0%.

The story came from my company.

The intellectual framework came from my work.

The conclusion reflected my values.

The examples reflected years of observing how people compensate for systems that do not work.

I conceived the article, drafted it, directed its development, refined its organization, read and reread the tone, and I edited it.

I approved every word published under my name.

But the machine found no human.

Perhaps it was looking in the wrong place.

The humanity of writing does not reside only in the statistical predictability of its sentence construction.

It resides in why the work exists.

• What it means.

• What experience formed it.

• What the author is trying to tell the reader.

• What she chooses to claim.

• What she is willing to defend.

And what responsibility she accepts when she presses Publish.

Pangram can assess the text.

It cannot assess the life behind it.

I’m NOT doing it. I’m choosing violence.

I will not deliberately make my prose worse to appear more human.

I will not remove my cadence.

I will not abandon parallel construction.

I will not stop using short paragraphs, repetition, humor, or clear transitions because a classification model has learned to associate polished patterns with machines.

I will not misspell words to prove I exist.

I will not put metaphorical fingerprints on every paragraph for algorithmic reassurance.

And I will not pretend I do not use AI assistance.

I do.

The work is still mine. The ideas are still mine. The voice is still mine. The judgment is still mine. The accountability is unquestionably mine.

Substack’s tool says I am 100% artificial.

That result will be news to the woman who has spent thirteen years building the body of work it just reduced to a bar graph.

Ahem.

Human: present.