Substack Writers: Stop Defending Yourself

The tool everyone is using to accuse you can’t clear you. That’s not a bug you can argue your way out of–it’s the whole design.

Substack Writers: Stop Defending Yourself

The tool everyone is using to accuse you can’t clear you. That’s not a bug you can argue your way out of–it’s the whole design.

The Jack Hopkins Now Newsletter #998: August, 14th, 2026

I have watched good writers on this platform humiliate themselves for three long G’ddamn weeks running (and it doesn’t appear to be done yet) and…not one of them did anything wrong.

Here’s the ritual.

Somebody runs a post through Pangram. The number comes back uglier than Mitch McConnell. It gets posted in the comments. And…then…the writer… a competent adult with a real readership…starts explaining.

“Here’s my process. Here’s my outline. Here’s a screenshot of my drafts folder. I only used it for research. I only used it for grammar. I swear I wrote every word.”

Stop it.

Not because you’re guilty. Because you are trying to win a proceeding that has no acquittal in it. And…once you understand the actual math…you’ll never post that silly-ass screenshot again.

The Number Nobody Reads

Everyone quotes the same statistic. “Pangram flags human writing as AI about one time in ten thousand.”

Independent work from the University of Chicago backs it up…nearly zero false positives across roughly 3,000 sample texts.

I’m not here to tell you that’s fake.

It isn’t.

Pangram is the best tool in this category…and pretending otherwise gets your argument dismantled in one reply. BAM.

Here’s the number nobody quotes.

Pangram’s own CEO told The Atlantic the false-negative rate is closer to one in seventy.

Sit with that.

Roughly one in ten thousand chance of wrongly calling a human a machine.

Roughly one in seventy chance of wrongly calling a machine a human.

That’s a gap of more than two orders of magnitude…and it points in exactly one direction.

The tool is built to accuse. It was never built to exonerate.

It Gets Worse

The Atlantic’s reporter ran the obvious experiment.

He had ChatGPT and Claude write articles. He ran them through a consumer “humanizer” tool…that shuffles wording and sprinkles in some grammatical roughness.

Then he fed the results to Pangram.

Every single time, Pangram called the laundered machine text…human.

So…let’s assemble the actual state of play.

A person who cheats…and spends four dollars on a humanizer: passes clean.

A person who wrote every word herself…in her own voice…on a Sunday morning: rolls the dice.

And…when she comes back with a clean score to defend herself…that score is worth almost nothing…because the false-negative rate means “human” is the answer the tool gives when it doesn’t know.

You cannot be cleared by a machine that clears everybody.

That’s the trap. It has no exit. And…every minute you spend posting evidence into it …is a minute you’ve donated to people who came to hurt you.

And “AI-Assisted” Means Anything At All

Now the category problem, which is worse than the math.

Pangram sorts text into generated…assisted…and human. The assisted bucket…per the reporting…can cover essentially anything short of copy-pasting from a chatbot… using it for research…having it argue the other side…using it as a thesaurus…running a grammar check.

Grammar check.

The company’s CEO says the model’s inner workings are largely uninterpretable…and that he isn’t certain how much more granular that “assisted” label can be made.

Read that again…because it’s the ballgame.

The tool cannot explain its own verdict… and the category it’s placing you in has no agreed meaning.

So…what is the accusation, exactly? That you used a spell-checker? That you asked something to steelman your argument?

Nobody knows.

Which is precisely why the label is such an effective weapon; it can be stretched to fit whatever the accuser already decided about you.

The Wall Street Journal’s James Taranto called it a ”defamation machine” after three op-eds got flagged at his paper.

Two of those writers acknowledged using AI to revise…and Taranto’s point stands: revising is not generating…and collapsing the two is unfair to the people who did the writing.

This Movie Already Played

If any of this feels familiar, it should.

Two years ago it was plagiarism-detection software. Activists ran academics’ work through algorithms…published the hits…and built a scandal.

A Harvard president resigned. Plenty of the accusations turned out to be thin…resting on tools that weren’t up to the job they were being used for.

Same structure. A machine produces a number. A partisan publishes it. The accused…is handed the impossible task…of proving a negative.

The difference…this time…is that the button sits directly under your post…and anybody can press it.

Now the Part That Actually Costs You Money

Everything above is why you shouldn’t have to defend yourself.

This is why defending yourself is bad business…and it’s the part I care about.

Look at who is demanding the explanation.

It’s never your paid subscribers. Not once…in any of these pile-ons…have I seen a longtime paying reader turn up to interrogate a writer’s process.

It’s drive-by accounts. Rival newsletters. People who never read you…and never will… who found a button that lets them play prosecutor for free.

Those people are not your customers. They were never going to be.

And…here’s the iron law…older than the internet and unchanged by it:

You cannot serve your buyers…and your critics at the same time. You cannot. Every hour spent on the second group…is stolen from the first.

When you post the drafts screenshot…three things happen, all bad.

You accept the frame; you agree that a number generated by a stranger’s browser extension is a charge requiring an answer.

You teach your own audience…that your work is suspect until cleared.

And…you hand a Tuesday…that belonged to your readers…to people who came to take it.

Defensiveness reads as guilt. Always. Especially when you’re innocent…because the innocent over-explain.

What To Do Instead

Not silence. Silence looks like hiding.

Do it once, in public…on your own terms…before anybody asks.

Two to four sentences: what you use, where the reporting comes from, who chooses the sources, who makes the argument…who decides what gets spiked.

Then…never litigate it again.

Somebody posts a score? “My standards are here. Same as last month.”

That’s the whole response. No screenshots. No proving a negative…to somebody who isn’t even reading or listening.

How I work.

For about four months now…virtually all my research runs through AI. I search at least two; usually ChatGPT and Claude…others if I need them. Between those two…I generally find what I’m after.

Then…I decide. Which sources I read. Which I keep. Which I throw out. And…every argument in every piece is mine.

So why do I still write the articles myself?

Because of what I call word fuckery.

If you’ve read me for a while…you already know it. The ellipses that break where they shouldn’t. The fragments. The sentences that would never survive an English class.

That’s not sloppiness. It’s pacing…built so the thing lands below the level you’re consciously reading at…and stays there. Recall is the whole point. Not comprehension in the moment. Retrieval next week.

If you’re not going to have access to the main points of the article…a week later…why read it in the first place?

AI is no match for me at that. Not yet anyway. But…I’ve been writing and speaking that way for over thirty-years.

I’m not writing for a grade in red ink in the corner of the page. I’m writing…so a good idea survives in your head after you’ve closed the tab.

One practical thing…and this comes straight from a writer who got falsely accused and beat it: keep your edit history.

Taylor Lorenz was accused of using AI on a Vanity Fair piece…Pangram’s own CEO investigated…and the tool had gotten it wrong. What she credited afterward wasn’t a counter-score. It was her revision trail.

Draft where your work is timestamped. Not to publish it…to have it.

What You’re Actually Selling

Your readers don’t pay you for keystrokes. They never did.

They pay for judgment. Which story matters this week. What’s actually true. Which detail changes the picture.

Which claim you refused to publish…because you couldn’t verify it…the one nobody will ever see…that cost you two hours and a damn good headline.

No detector measures that. It isn’t in the surface features. It never was.

The one running Pangram on your post is measuring the only thing a machine can see…and calling it the thing that matters.

Don’t help him.

If you’re a Substack writer…publish your standard once. Then get your ass back to work…and stop explaining yourself like a child does to every adult who resembles a high school principle.

#HoldFast

-Jack

Jack Hopkins

P.S. Ask your accuser one question: what score would satisfy you?

He won’t have an answer. There isn’t one…a clean score means almost nothing…and he knows it…or would…if he’d read past the headline.

That’s how you know it was never about the number.

Oh…and…if you’re a membership subscriber…stay tuned for today’s Bedrock series article. It’ll be out in a couple hours…or sooner!

Sources

• The AI-detection arms race is a disaster in the making — Matteo Wong, The Atlantic, May 2026. Source for the one-in-seventy false-negative rate, the humanizer test, the Lorenz and Taranto episodes, and Max Spero’s comments.

• Pangram AI detector: is Substack’s new scanner accurate? — on Substack’s July 21 rollout and what the displayed percentage does and doesn’t represent.