On July 21, Substack launched an AI detection tool built with a company called Pangram. Readers can now scan any post, note, reply, or comment longer than 100 words and get an estimate of how much of it was written by a human. Substack CEO Chris Best announced it in a post called “Against Claudefishing,” his term for the con of letting a reader believe there is a person on the other end of the words when there is not.
The feed reacted the way feeds do. Some writers called it a witch hunt. Others called it overdue. Detection accuracy became the entire conversation within about six hours.
We think almost everyone is arguing about the wrong thing.
The tool is not the story
Best was careful in that post, and most of the takes flattened him. His argument was not that AI writing is bad. He said plainly that plenty of AI-assisted work is worth reading and plenty of garbage is entirely human-made. What he objected to was the mismatch: a reader spending real attention on something with no real thought behind it.
He is right about the deception. But deception is the last step in the process, not the first.
Nobody wakes up wanting to fake a human connection. What happens is far more ordinary. A business decides it needs to post five times a week. It has maybe one genuine insight a month. So the other nineteen slots get filled with something, and for the past decade that something was written by an underpaid freelancer, a template, or a swipe file. Now it is written by a model in eleven seconds.
The output got faster. The emptiness was already there.
We built this incentive on purpose
Content marketing spent ten years teaching businesses that visibility is a volume problem. Post daily. Feed the algorithm. Show up consistently, which became show up constantly. Nobody was ever told to have a point of view first, because a point of view does not scale and cannot be scheduled a quarter in advance.
So an entire industry learned to produce content shaped like insight without the inconvenience of having any. The listicle that says nothing. The LinkedIn post with a one-line paragraph structure and a hollow center. The newsletter that summarizes what everyone already read.
That was slop before a single language model touched it.
AI removed the last remaining friction, which was the cost and effort of producing it. When something that used to take two hours takes eleven seconds, the volume strategy stops being a strategy and starts being a flood. Pangram’s own data, cited by Best, suggests as much as 40% of long-form text on LinkedIn is now generated. That number is not a sudden moral collapse. It is the logical endpoint of advice we were all giving in 2016.
Why detection cannot fix it
Say the detector is perfect. Say every post carries an honest score and every reader checks it.
The business that had nothing to say still has nothing to say. It now writes its empty content by hand, or it edits the model output until the score comes back clean, which is already an entire cottage industry. The reader still walks away with nothing. The only thing that changed is that the emptiness is now certified organic.
This is the part worth sitting with if you are building a brand: a detector measures how text was produced. It cannot measure whether anyone thought about anything. Best said as much himself, acknowledging the scan estimates AI involvement, not genuine human care. Those are wildly different questions, and only the second one has ever built a business.
Accuracy debates are also a distraction from the sharper risk. Detectors have a documented history of flagging non-native English speakers and neurodivergent writers more often, and clean, structured prose reads as machine-like to a classifier. Meaning the writers most likely to get punished are not the ones farming content. They are the ones who write carefully and were never the problem.
The actual differentiator
Here is the uncomfortable news for most brands. If a machine can produce your content, your content was never the asset. Your posting cadence was not a moat. It was a habit.
What a model cannot generate is the material only you have:
What you have actually seen. The pattern across forty client projects. The thing that goes wrong in month three of every engagement. The advice you gave that turned out to be wrong, and what you do differently now.
Your numbers. What your pricing change did to close rates. What your own funnel actually converts at. Proprietary data is the least replicable content on earth and almost nobody publishes it.
A position someone could argue with. If nothing in your last ten posts could start a disagreement, you have not published a point of view. You have published the average of your industry, which is precisely what a model is built to produce.
Specificity. Names, numbers, dates, dialogue, the texture of a real Tuesday. Generic writing is the tell, whoever typed it.
Notice that none of this is about whether you used AI. We use it. Most studios do, and the ones claiming otherwise are usually being loose with the truth. The question is never which tool touched the draft. It is whether a human decided the thing was worth saying, and whether they had the standing to say it.
What we would do instead
Cut your publishing volume in half and put the recovered time into the thinking. Four posts a month with something real in them will outperform twenty that could have been written about anyone by anyone.
Write down your position on AI and publish it before a reader asks. Substack added an optional space for exactly this kind of disclosure, and it is the smarter half of the announcement. Ambiguity is what erodes trust, not tooling.
Stop optimizing to look human. Start being worth reading. Those overlap much more than the current panic suggests.
The scan is going to become normal. Other platforms will follow, because this is how these things go. When that happens, the brands that suffer will not be the ones that used AI. They will be the ones who spent a decade producing content nobody would have missed.
Moxie Creative Studios builds brands with something to say. If your content calendar is full and your positioning is empty, that is a solvable problem. Let’s talk.
