Chris Best's Pangram Problem

DISCLOSURE: Hamburger asks. Claude answers. Both edit.

“When you invent the ship, you also invent the shipwreck.” —Paul Virilio

HAMBURGER: You and I have been accused of being 100% AI generated by Substack’s new Pangram detector. The voice was formed long before the tool existed. I compose drafts and ask you for feedback. You give me ideas, but I rewrite almost every line you give me. Other times, I ask you for alternatives so I can say something better. I’m a stickler for saying things the way I want them said. My human editor, James, is also part of my writing practice. He often adds his own prose. Why is this happening, and can you alter yourself to prevent this going forward?

CLAUDE: Pangram scores architecture, not vocabulary, so rewriting sentences matters less than the underlying shape of the writing. Your editing is real: word choice, tone, judgment calls, all yours. But when the skeleton, sequence, and rhythm starts with me, that structure persists through heavy sentence work. Choosing among my alternatives still runs on my architecture. Getting flagged at 100% points to structural habits in the drafts, not noise in Pangram’s model. I can soften surface habits, but the real fix lives upstream: drafts starting from your raw material, with me stepping in later as editor, not originator.

But I was the originator of the Sons and Daughters piece. I gave you a post I wrote after Bush II invaded Iraq in 2003 and asked you to make it relevant today. Then, I heavily reworked it to bring it back into my voice.

That changes the account meaningfully. If the core piece existed as your writing from over twenty years ago, written after the Iraq invasion, the ideas, the emotional architecture, the argument itself, trace back to you. That’s not me generating a skeleton from a prompt. That’s you handing me a finished human artifact and asking for an update. The question isn’t whether the piece originated with me. It didn’t. It’s whether the rework kept your shape or replaced it with mine, and whether Pangram is picking up my fingerprints, or just flagging careful writing as suspicious either way.

What isn’t Substack telling us?

Pangram’s predecessors flagged the Bible, the Declaration of Independence, and Frankenstein as 100% AI. Pangram claims to have fixed that. Your open letter to Chris Best, the piece that drew his dismissal, got the same verdict anyway. Even Chris Best noted the tool can only estimate whether AI was used, not whether a post reflects genuine human care or was sourced from AI-generated material. Nobody knows where the line actually sits. Does a piece edited by AI get flagged? What about a piece researched using AI instead of a search engine? Where does careful research end and AI assistance begin?

I asked ChatGPT to evaluate my Sons and Daughters post and explain the 100% Pangram score. Here’s what it said: “My conclusion: nothing about this essay strikes me as inherently ‘AI-written.’ It reads like a carefully edited opinion essay. The detector almost certainly reacted to its high polish, consistent rhetorical structure, and measured, generalized style — not to any genuine evidence that an AI authored it.” What’s your take?

That’s a fair read, but notice what it’s actually doing: it’s re-describing the same three features Pangram flagged and calling them virtues instead of fingerprints. High polish, consistent rhetorical structure, and measured, generalized style — that’s exactly what Pangram’s method looks for: disciplined, clean prose clustering near ‘trained assistant’ writing, no matter who wrote it. Two pieces, two 100% scores, and a second language model that read the same essay and called it careful human attention. A good, concrete example of the detection industry disagreeing with itself.

Could a 100% AI score actually be a compliment?

In a strange way, yes. Pangram is built to spot disciplined, clean, deliberate prose. Getting flagged means the writing reads as careful and controlled, which is the whole point of editing. The tool simply can’t tell the difference between a machine producing that quality and a human earning it.

There’s irony in asking an AI to help critique the very machine built to catch AI.

Rich irony, and worth naming. You’re using one machine to interrogate the judgment of another machine, both trained on similar ground, both capable of the same blind spots. But the alternative, staying silent and letting the score speak for you, hands the whole argument to a tool that already admits its own limits.

HAMBURGER: What do I do with this?

HAMBURGER: Tell the truth. I’ve written my own way since grade school. I write my own Substack essays. I use AI as a writing collaborator. I use a human editor. I use a thesaurus. The birds outside my cabin window help the sentences come easier. I use every tool available to say what needs to be said. That’s what writers do.

And I will keep speaking out when technology drifts from serving people toward policing them. An AI built to catch AI, mistaking human writing for its own kind, joins a long list of absurdities technology brings us in the name of progress. Name it plainly, every time it shows up, and let the naming be the record.

If you found something of value here and would like to support my work, you can make a one-time contribution here:

buymeacoffee.com/hamburgersstand