Substack's AI Detector Said 80% AI. Does the Thinking Still Count?

About 2,400 years ago, Plato wrote down a story about the invention of writing. A king hears the inventor's pitch and isn't impressed. He says writing offers people "the appearance of wisdom, not true wisdom."

Replace the word writing with AI, and you have nearly every argument I've heard this year.

The argument never really went away. Editors cut and rearrange authors' books. Ghostwriters write the memoirs of famous people. Speechwriters put words in presidents' mouths. Over time, we mostly made peace with all of them.

What the scanner measured

In July, Substack added an AI detector, powered by a company called Pangram. Readers can see an estimate of how much of a post was written by a human and how much was "AI-assisted." Substack says the goal is openness about process, not punishment.

By the tests we have, Pangram is good at its job. But look at what a 2025 University of Chicago comparison tested: work written entirely by a person against work written entirely by a machine. It didn't test what many writers who use AI actually do, which is a person and a tool working together on the same piece.

So what does a score mean when the honest answer is "both"?

When it landed on my own work

Most of my essays came back around 80% AI. A number like that doesn't come with an explanation. It just sits there next to your name.

The full essay explains how each piece gets made, where I draw the line, and the part of this story that I don't love. The larger argument about who does the thinking, and what we owe the people whose work built the tools, is the kind of thread I pull on in my book, Pulling the Thread.

The question I haven't been able to settle: if a tool helps you say what you actually mean, but it was built partly on other people's work without their permission, is using it openly enough?