I notice the exact moment it happens. I’ll be half-watching a YouTube video, the narration hits one of the tells (you know the ones — that em-dash thing, the seesaw parallelism, the whole “this isn’t X, it’s Y” move), and I check out. Instantly. Almost involuntarily. There’s a whole taxonomy of these patterns now. I can never keep the names straight, but thankfully (maybe?) my brain does it for me.
And the weird part: there’s a person right there on camera, saying the words in their own voice, and it still sounds like no one. The ideas are usually fine. I’m tuning out material that is, sentence by sentence, perfectly competent...
What I’m actually scanning for is proof of thought. Some evidence a human mind passed through the material on its way to me. More and more, there is none.
For basically all of history, text was a reliable proxy for thought because text was expensive. A four-page memo meant somebody burned an afternoon on it. Length was the peacock’s tail. It was expensive, so it meant something. Now producing words costs nothing, so the words certify.... nothing.
But reading costs what it always did. Attention is still the same size; there’s just more and more chasing it, which I know isn’t new. And these days, my attention is at a premium. So when someone sends me “raw model output” or I come across it, they saved an X hours of compression and billed it to my attention. The work of turning information into meaning still has to happen somewhere. Now it happens at my end of the table? Fine - I’ve now decided I no longer value this information…
There’s an old line: I would have written a shorter letter, but I didn’t have the time. Working with writers most of my career, I understand better than most that shortening was always hard.
That’s why unedited AI output reads as disrespect. Frankly, that’s the word for it. Decoded, the message says my hour matters more than yours. Or I couldn’t be bothered. I’m now a meat proxy, as my cofounder and i endearingly call ourselves...
The obvious objection (according to my lil ai homies) is that I’m pattern-matching on style, and the patterns are temporary. Sure... in a year the cadence detector in my head will be useless.
So what counts as proof of thought?
Compression, for one. You can only summarize what you understand, which is what makes a genuinely short email so nice.
Specificity: the detail only someone inside the situation could know.
When I started writing this I thought the strongest signal was a decision. Options cut. A position taken. Something the writer could turn out to be wrong about.
Maybe it’s accountability?
Maybe it’s typos, the imperfections?
I don’t claim to know the answer. I don’t claim to be a good writer, either. You’re still reading this? People still read?
For now, I know the sound of “no one home.” But who needs to be home when all our machines are just talking to each other? That’s gonna have to be another essay...
PS: The first draft of this essay was flagged as 100% AI-generated. This version scores 92%. The detector is literally pattern-matching on style which is exactly what I predict goes stale in a year.