What Percent of This Is You?
Everything's AI-written now. The question is what fraction of the thinking is yours.
Scorecard, applying the thing this post proposes:
| Claude | Ben | |
|---|---|---|
| Written (by characters) | 93% | 7% |
| Ideas | 15% | 85% |
I wrote in February that I’d spent two years treating all LLM writing as slop, and that the real line was between using AI as a microwave burrito and using it as a compiler for your own thinking. I still believe that. But I left something out, and it’s been bugging me for months: tone and substance are two completely different problems, and I mashed them into one post.
Let me split them apart.
Tone first, because it’s the easy one.
Sometimes I read something obviously LLM-written and I like it. Not “tolerate.” Like. But look closer at why: it’s usually a chat window that’s been mimicking my tone back at me for an hour, adapting word choice and rhythm until it’s basically doing an impression of me, to me. Of course I like it. Duh. I’d like anything that flatters my own cadence back at me. The catch is that the thing I’m reacting to isn’t quality, it’s personalization, and personalization doesn’t transfer. Hand that same paragraph to Jeremy and he has no reason to feel anything about it. The writing isn’t secretly good. I’m just the only person in the world it was built to please.
Then there’s the other kind. LinkedIn thought-leadership posts with the single-sentence paragraphs. The “here’s what nobody’s talking about” openers. The em-dash doing the emotional labor a real transition should be doing. I hate that stuff, and I hate it specifically, the way you hate the 1-877-Kars4Kids jingle, not writing in general.
So: tone is real, it’s noticeable, and it can cut either way depending on whose voice got trained into the thing. But it’s a side issue. It’s not what actually decides whether something is worth my time.
The actual problem is thinking, not tone.
Here’s my prior when someone hands me a memo, a PRD, a blog post, an email that reads like it came out of ChatGPT with zero friction: there’s a two-sentence prompt buried somewhere in this document’s ancestry, and everything else is output tokens. Ninety-nine percent autocomplete, one percent intent. And reading that is a shit use of my time, because I’m not getting access to a person’s judgment. I’m getting access to a language model’s median guess at what a person’s judgment might sound like, with the actual person having contributed almost nothing except the decision to hit enter.
That’s the thing that makes me allergic — the absence of anyone home behind the prose.
Now the counterexample, which is basically all of my own writing.
I obv write almost none of my own words by hand anymore. I core-dump thoughts into Claude, argue with the draft, paste in feedback from a second model, argue with that too, rewrite, cut, rewrite again. By the time something ships, I might have typed two hundred words of the four thousand that make up the post. The rest came out of a model. Natch.
And yet I’d bet real money you can’t find the seam. (Sometimes you can, but that’s usually on purpose — see the Cinderhall post, where the AI-ness is loud and stupid and part of the joke, not something I was trying to hide.) Not because I’m some prompt-engineering savant, but because every sentence in there passed through my judgment at least three times before it shipped. The words aren’t mine. The decisions about which words survive are.
As are the cultural references that are so simultaneously low brow and high brow that no one but Kenny, Jacob Tierney, and Jared Keeso appreciate them. That’s what I actually mean by the content of a piece of writing — who decided, not who typed.
So the question I actually want answered isn’t “did AI write this.”
Everything’s AI-written now, in the trivial sense. The question is what fraction of the thinking in front of me is the author’s, versus the model filling gaps the author never bothered to close.
And I don’t mean that as a token count — a document can be ninety percent human-typed and still be entirely the model’s idea, if the human was just transcribing a summary they half-understood. (LOL wut? Yeah, that happens more than people admit.) I mean it as a measure of where the judgment lived.
I want a number for that. Something like: 80% author, 20% model — meaning the ideas, the structure, the specific claims, the stuff that would be wrong if you cut a paragraph, all came from a person who sat with the problem. Versus 5% author, 95% model — meaning someone typed a sentence and a half and let the gaps fill themselves in, and what you’re reading is mostly the model’s extrapolation of what a competent person in that role might say.
This isn’t new, it’s just newly urgent.
Ghostwriting has worked this way forever. A public figure has maybe ten pages of genuinely interesting material — a real insight, hard-won, specific. A ghostwriter turns that into two hundred and fifty pages because that’s what a book contract requires, which is kind of bullshit but also just the job. And readers have always had a version of this radar. You get two chapters in and you can feel whether there’s a person underneath the prose or just a very competent professional stretching ten pages into a bestseller-shaped object. The ghostwriting industry is basically a case study in how much bulk you can add to a small kernel of actual thought without most readers noticing until they’re already three chapters deep.
LLMs just made that trick available to everyone, instantly, for free, at the bottom of every text box. Which means the ten-pages-into-two-hundred-fifty problem isn’t confined to celebrity memoirs anymore. It’s your coworker’s product spec. It’s the “quick thoughts” email that’s actually four paragraphs of a model padding out one real sentence. It’s most of LinkedIn.
So here’s the actual proposal.
I already do a version of this informally — I’ve posted the raw prompt behind a piece before, specifically so people could see the ratio for themselves. I want to make that a habit instead of a novelty. A line at the top of anything I publish: author-thinking versus model-filled-in, my best guess, no more scientific than that. I doubt many people will follow. I don’t really care, natch. If more people told me upfront how much of the ten pages was actually theirs, I’d read a lot less garbage, and I’d stop skipping the good stuff that happens to look like it.
I'm Ben Stein, co-founder and CEO of SuperDuper, helping underwater parents manage all their family logistics. I live in Oakland with my Keeper wife Arin, our two overprogrammed teenage boys, and a dog named Soup who contributes nothing helpful to the family logistics and arguably makes them harder, but he's really cute.
If you want to follow along: superduperlabs.com · benjaminste.in · LinkedIn · Substack