AI Detection and the Presumption of Guilt

Also watch and listen

The signals

• Jonathan: the label does its work whether or not it is true. Substack shipped a button that reports whether a post is AI generated, AI assisted, or human written, and his position is that arguing about its accuracy concedes the wrong ground. His anchor is the nocebo case: New York announced fluoridation, the rollout ran a month and a half late, nobody told the public, and on the day it was supposed to start people called in with headaches and stomach aches. Nothing was in the water. What changed was what people believed about it. The second anchor is a Monet posted as AI art, panned as soulless by a crowd that deleted its replies once the painter was named. He ran the same image past ChatGPT and Claude with the same framing and got the same verdict, which means the bias is not only human. A detector does not need to be wrong to do damage; it only needs to be believed.
• Sam: origin and quality are different properties, and only one of them is being measured. Textual origin, production process, and communicative quality are three separate things. Pangram estimates the first from a finished artifact. Disclosure lets an author testify about the second. Nothing on the platform touches the third. Meanwhile the cost of producing plausible language collapsed and the methods for evaluating meaning did not move. Knowing an apple came from a local farm without pesticides tells you what the apple is; it does not tell you whether it is a good apple. When someone asks how an article was written, the answer worth wanting is "very well, thank you."

The subtractions

• Jonathan: subtract the AI-assisted label, everywhere, including the voluntary kind. Not only Substack's machine-applied one. He held this position before the detector existed and holds it against self-disclosure too, on the grounds that journalism has never owed its audience its sources, its editors, or its ghostwriters, and that volunteering the information is what teaches an audience there was something to be suspicious of. His line on the scan button: you are not entitled to my process. His last word on tape: engage with my ideas, not my process.
• Sam: retire authorship as a quality claim, in both directions. Human-written is not a quality claim and machine-flagged is not one either. Both are proxies, and cheap ones, reached for because they are easy. What is newly possible is naming what good actually means and measuring that directly, which is the thing neither the label nor the detector attempts. This is a genuine disagreement rather than the same subtraction twice: Jonathan removes the signal, Sam keeps the signal and strips the quality meaning out of it. Neither talked the other out of it on air.

About this episode

Two voices, no cohost, and two recordings. The conversation was livestreamed on Wednesday while Limited Edition Jonathan was in a car on a two-hour drive to the emergency room where his son had been taken from summer camp. He knew that when the session started and kept the commitment anyway. The cell connection did not survive the trip. The episode is that recording, with host pickups filling the places where the connection took an answer mid-sentence, and the drops are left audible rather than smoothed. It is his argument with commentary, not a two-hander, and it is the first episode of the show assembled this way.

The clearest casualty is the question at 7:39: how does this actually hurt Substack and Substack readers? He gets one sentence in before the line goes. The answer, from his published piece: the flagged writer stops publishing. Not a debate, not a correction, just a vacuum. Stigma does not distribute by merit, it lands on whoever is least equipped to fight it, and his whole case for AI is that it lets a writer without a team compete against people who have one. Those are the writers a label discourages first.

The asymmetry is the part neither the tape nor most of the coverage reaches. The scan runs downward only: writers are scanned, the platform's own posts are not, and Chris Best's announcement of the feature carries no disclosure of its own. Writers are not notified, and the appeal arrives after the label already exists. A presumption of guilt with a customer service form attached.

Worth stating what Best actually argued, since the episode spends its length disagreeing with him. His piece defines the problem as deception, text passed off as human when it is not, which he and Freddie deBoer both call a con. On that definition the target is the faker. What shipped scans everybody and flags assistance of any kind at a hundred words or more, so it reverses hardest on honest, disclosed, assisted writing, which is the category his definition was never about. Best writes that when he wants Claude's opinion he will ask Claude; the button is a machine's opinion about whether a writer is worth reading.

The show streams to Substack and this episode's Substack surface is part of how it ships, so the argument is being made in the room it is about. Named once on air, not litigated.

This week's links:

• Nothing Was In The Water — Limited Edition Jonathan's piece, the premise of the episode
• Against Claudefishing — Chris Best's case for the feature
• Writing Quality After AI — Sam's article on origin, process, and quality as separate properties
• Limited Edition Jonathan on Substack

No sponsor this week.

Want to bring your own signal and subtraction? Find yours.

The newsletter

Issue 62 · Sunday 2 August 2026 · sent on Substack and LinkedIn.

One signal, one subtraction, one analogy, in five minutes. It draws on this episode but stands on its own. Every issue.