Table of Contents
Do not index
Should you run your drafts through an AI detector before you publish? No. Editing your writing to satisfy a machine is the fastest way to lose the voice readers and clients actually pay for, and the CEO of one of the largest writing platforms on the internet just made the same argument in public. The detector debate is not really about catching bots. It is about whether writers start performing humanity for an algorithm, and that is a trap worth naming before your whole content operation walks into it.
The context is a platform fight worth watching. After Substack shipped reader-facing AI detection built with Pangram, Medium CEO Tony Stubblebine publicly rejected the approach. His argument, made on Medium, is that detectors only punish good actors while slop producers defeat them with humanizer tools, and that platforms should rely on human curation instead, human firewalls that judge wisdom and truthfulness rather than word choice. His warning to writers was direct. "If you give in to this Substack system, you're changing your own writing to make it less human just because the AI detector doesn't actually know any better."
We have run this experiment before. A decade of SEO-brain taught a generation of marketers to stuff keywords, pad word counts, and sand every edge off their prose because a machine was grading it. The content ranked and nobody read it. Writing to pass AI detectors is the same failure mode wearing a new badge. You end up swapping precise words for quirky ones, breaking rhythms that were working, and second-guessing clean sentences because a probability score said they looked synthetic. The machine gets pleased. The reader gets less.
This warning is for ghostwriters charging $5k to $30k per month whose product is a client's actual voice, for agency owners between $200k and $2M in revenue running content teams that ship daily, and for founders whose personal-brand content carries their pipeline. If you are still selling raw volume, 200 generated posts a month with no human judgment attached, skip this article. Detectors are not your problem, your product is, and no amount of humanizing will fix a model built on saying nothing.
What separates content that survives this era from content that does not is something I call the Judgment Layer. It is everything in a piece of writing that a detector cannot measure and a model cannot supply. Deciding what is true enough to stake your name on. Deciding what to leave out. Knowing which client story makes the argument land and which one muddies it. Knowing what your reader tried already and why it failed. Detectors score word choice. Readers score judgment. Every hour you spend gaming the first is an hour taken from the second, which is the only score that converts.
Why AI detectors punish the wrong writers
Stubblebine's structural point deserves more attention than it got. Slop producers are the best-equipped players to beat detection, because their entire operation is tooling. They will run output through humanizers until it passes, at scale, at near-zero cost. The writers most likely to get flagged and most likely to contort their prose in response are the ones writing in good faith, often in a second language or a plain professional register that pattern-matches to machine output. A detector-first culture asks the most human writers to act less like themselves while the least human operations sail through. That is backwards, and building your editorial process around a backwards incentive compounds the damage weekly.
The deeper problem is what optimization does to a voice over time. Voice is cumulative. Readers subscribe to the way you see things, expressed the way only you express it. Start editing toward what a classifier considers human and you are averaging yourself, one revision at a time, into the same beige middle the slop lives in. The moat fills itself in.
Where human judgment actually shows up in content
Run your operation on reader signal instead of machine signal. The Judgment Layer shows up in specifics no tool can generate. The number from last quarter you were slightly embarrassed to share. The position that cost you a prospect but won you the right client. The paragraph you cut because it was clever and useless. It also shows up in what you measure. Replies from real prospects, conversations that turn into deals, and retention of the audience that matters beat any dashboard score, which is the same argument I make about measuring LinkedIn success beyond the analytics dashboard. A detector score is just another vanity metric, and it is not even measuring your reader.
The strategic implication is worth sitting with. Platforms will spend the next few years fighting about detection, curation, and proof of humanity, and none of that fight is yours to win. Your trajectory depends on the one input every side of the debate ends up rewarding, which is judgment applied in public over time. The writers and agencies who keep compounding that asset will be cited, trusted, and hired regardless of which detection regime wins. The ones who spent this era rewriting themselves to fool a classifier will discover they optimized for a machine that was never going to pay them.
