The Written-by-a-Human Signal
The tell isn’t the writing quality anymore.
AI produces competent prose. Structured arguments. The kind of thing that would have been a solid B+ on most assignments. It covers the topic, hits the points, doesn’t contain any lies you’d immediately notice.
What it doesn’t contain is anything you couldn’t have found somewhere else.
What the model can’t generate
Every piece of writing I’ve read recently that didn’t feel like a content mill had one thing the AI version couldn’t: something the author had lost.
Not lost as in “failed and reflected on it from a comfortable distance.” Lost as in — here is the specific thing I got wrong, here is the specific cost, here is what I now know that I didn’t before. That texture — the embarrassing detail, the number you’d rather not publish, the name of the exact mistake — is what makes writing land.
A language model can produce a failure narrative. It can write “I launched a product that didn’t find product-market fit, and I learned the importance of customer discovery.” That is a failure story the way a Wikipedia summary is a biography. The scaffolding is there. The person isn’t.
Personal stakes aren’t optional
The difference between content that gets saved and content that gets scrolled past is usually this: does the writer have skin in the outcome they’re describing?
Not credentials. Not even experience. Skin. As in: this cost something, and the author is willing to say what.
That’s what creates the written-by-a-human signal — not a byline, not an author photo, not an elaborate bio about your decade in the industry. It’s the sentence where you clearly had something to lose and you put it in anyway.
A few things that carry that signal:
- The specific number you didn’t hit. Not “the campaign underperformed.” The actual return, the delta, the quarter where it all went sideways.
- The decision you’d now make differently. Not “I learned a lot from this experience.” The exact call, explained, with the constraint you didn’t see until after.
- The belief you publicly held that turned out to be wrong. The harder this is to write, the more it’s worth publishing.
None of these can be reverse-engineered from research. They require having been in the situation — which means they require a person.
The differentiator that compounds
The irony is that this has always been the differentiator. Specific, stakes-visible writing has always beaten generic authoritative prose. AI content didn’t create the problem — it just made the gap between the two types harder to ignore.
The sea of “high-quality” content is bigger now. So is the signal-to-noise gap. Which means the piece that lands is the one where you had a good reason to vague things up and chose not to.
The model can write anything except the part that happened to you.
That’s the only part worth reading.