Dear Chris Best and Substack Team
It’s morning on the homestead.
Coffee on the table, all notifications silent, the way I write everything.
I read your letter about Claudefishing and the new Pangram check inside the Substack app.
I nodded most of the way through.
Then I reached your last line, the one about asking Claude for Claude’s opinion, and I laughed, because I agree with you.
And then I opened my own draft.
I speak six languages at different levels. In two of them I could write a book. I have written one. None of them was English.
I think it first in Swiss German. A spoken language with no official way to be written down. You speak it. You don’t spell it.
From there I move it into literary German, then across into English. Not any English. American English, the one the tool was trained to trust.
And that word, American, is where I want to stay calm and still say the hard thing.
There is an assumption baked deep into your letter, and it runs deep in the American way of reading the world: that the world thinks in English, and in your English above all.
Let me assure you: It doesn’t.
Most of the planet spends its whole life translating itself into your language so you will listen.
We chose to do that. We adapt, because that is how you get heard. (!)
But when a tool then reads the seams of that translation and calls it fake, the message the rest of the world hears is old and familiar: your language is the only one that counts.
I don’t say that in anger.
I say it because that is exactly what the world sees when it looks at America, and I think you’d rather know than not.
So here is my problem with a feature I otherwise understand.
You named the right sin and shipped the wrong test
Claudefishing is real.
A reader who invests attention in words with no mind on the other end has been conned.
Freddie deBoer is right that fooling someone that way is a con. You are right to want the human protected.
I want that too.
But read your own sentence back.
You wrote that Pangram “can only detect whether AI was used to make the text, not whether great human care went into creating it.”
That is the whole thing, right there, in your own words.
You want to protect the human behind the writing.
The tool you shipped measures which keyboard did the typing.
For a reader in Brooklyn who thinks in English and writes in English, those two are the same thing. For me, and for most of the planet, they split apart.
The mind behind my sentence is mine.
It thought the thought in a language Pangram has never read, because nobody writes that language down.
The English is the last step, not the first. When the check flags my paragraph, it is not catching an absent human.
It is catching an accent.
A teaspoon of soil holds more than the field shows
More life moves through one teaspoon of healthy soil than walks across the whole field above it.
You judge the field by what you can dig into, not by the surface. Most of the work is underneath, invisible, and it is still the realest part.
My thinking happens under the surface of the English sentence, in the language the tool can’t dig into.
Pangram reads the topsoil and declares the ground dead.
The care Chris Best says the tool cannot see is exactly the care I put in. It just went in one layer down, in German, before the English ever existed…
The trust signal renames itself
Here is the distinction I think you missed, and it is a factual one, not a rhetorical one.
You built this as a trust signal: is there a human here or not?
But the variable it actually measures is fluency, not presence. A study out of Stanford in 2023 (Liang and colleagues, in the journal Patterns) ran seven detectors against real human writing.
On US eighth-grade essays, near perfect.
On essays by non-native English speakers, more than half were flagged as machine-written. Same tools. Same question. The only thing that changed was where the writer learned English… (!)
I can’t tell you if Pangram behaves exactly like the tools in that study.
You say the research on it is strong, and it may well be.
What I can tell you is that the incentive which produced that bias has not gone anywhere. A detector rewards writing that looks the way native, unassisted English looks.
My honest writing does not look like that, and it never will, because I did not grow up inside “your” language.
So the signal quiets me twice.
Once because I use AI to reach clean English, which you say is fine, and once because the very cleanness gets read as absence.
Chris, that is not protecting authorship. That is protecting the writers who were born on the right side of a language… (!!)
What I’d actually ask for
You opened the floor to debate, so I’ll take you up on it.
Keep the check. Keep the “How I make this” statement; that one is good.
But show the reader the honest version of what the number means.
Not “how human is this,” which the tool cannot know.
Something closer to “how much does this resemble unassisted native English,” which is all it can measure. Those are different questions, and a reader deserves to know which one they are being handed.
And overall.
Trust the reader more than the meter.
If a piece serves them, moves them, tells them something true, they mostly do not care which tool touched the grammar.
They care whether a person meant it. I mean this. I thought it in a language you have never read, and I stand behind every word after it crossed into yours.
When you want Claude’s opinion, ask Claude.
Fair.
But when you built a tool to find the humans, you built one that can’t see the ones who had to travel the farthest to reach your page.
What are you going to do about us?
Let’s regenerate the world - starting with yours,
Daniel
Exactly my point. Love Greek. In younger years I've worked in Crete.
Thank you Kevin for supporting my claim. Our world is rarely painted in black or white. And the use of AI is clearly more gray... The mentioned study in the article actually proves that AI detection doesn't work at all...