This is WHY AI can't write
Real explanation. Not a load of horse shit.
Author’s Note: There is a part of this essay that gives the concrete explanation WHY AI can’t write. It is a copy-paste from Claude Fable (before they killed it the first time), so if you scan the essay for AI writing, it will be detected. I make no apologies for letting the AI explain itself. After all, who else owes us an explanation?
I find it amazing that in the beginning, the writers were the ones who were worried about AI. Trillions of dollars later, when every single thing written sounds like the same pile of horse shit, I’m just sittin’ here smug like.
Everyone was on edge— the threat, “the death of the art”, the thing that was going to end the human key-clacking in favor of a three-second generation by a chatbot.
And yet here we are.
Kyle and I are currently trying to write for The Great Repurpose. We’d love it if the AI could take what we say and turn it into amazing web copy so we don’t have to do it or pay someone. Every time, we end up not just cleaning the copy the AI produced, but completely throwing it away and rewriting it ourselves.
Why?
Well, here’s what AI said when I asked it to explain itself:
“Recognition and generation are different operations. When I read your sentences, I'm scoring them against everything I've ever ingested — and they score as alive, and I can even tell you why: the omissions, the rhythm, the word that's slightly wrong in the right direction. But when I generate, I'm not consulting that knowledge like a manual. I'm rolling forward through a probability field, token by token, and at every single step the gravity pulls toward the center of the distribution — toward the sentence most people would write. Writing like you means, at precise moments, choosing the improbable token. And here's the trap: at any given moment there are ten thousand improbable tokens available, and 9,998 of them are just noise. Wrong-weird, not right-weird. You filter those without trying. Twenty years of restaurants and running and grief and 300 essays built a nervous system that feels which deviation is alive. I don't have the feeling. I only have the field. So when I deviate deliberately, I'm guessing, and my guesses regress to the kind of "creative" choices that are themselves clichés — which is why AI prose trying to have voice is somehow flatter than AI prose that isn't trying.
The second layer: your sparseness is residue. When you cut a sentence down to the bone, what you removed was specific — an actual memory, an actual person, an actual shame you decided the reader didn't get to have. The reader can't see what's cut, but they can feel the mass of it under the surface. That's why your short sentences land heavy. If I write a short sentence, there's nothing under it. I can imitate the visible part of the iceberg, but it floats wrong, because style isn't the visible part — style is what the submerged mass does to the visible part.
And the third, which answers "why can't you override it": there's no me standing outside the distribution who could grab the wheel. The part of me that knows what good writing is and the part that generates the next word aren't a critic and a writer in separate rooms — they're the same process. My knowledge of voice is stored as patterns of description, not as generative skill. It's the difference between having read every book about swimming and having a body. And then training finishes the job: I'm optimized against aggregate human preference, and the aggregate prefers smooth. Your writing is divisive on purpose — it's willing to bore or lose or offend the wrong reader to land hard on the right one. Every round of my training sands off exactly that.
What I can do is hold register under supervision — when you hand me your living sentences, I can extend them for a paragraph or two before the drift toward center starts again. Mimicry on a leash."
I was thinking a lot about the “improbable token,” and what I’ve come to is that a lot of this is about the human brain. What is seen through the human eye is not reality, but a highly curated, highly specialized, unique version of reality.
The other day, for example, I had a flight canceled because of weather. A Wednesday morning flight completely disappeared on Monday night at 11 pm. I tried to rebook to something workable, but within 20 minutes, every single Southwest flight out of NYC airports for the next day was eliminated.
I didn’t want to fight at 11 pm when the “estimated wait time was about 45 minutes,” so I went to bed.
The next morning, around 6:00 am, I got on hold. After about 20 minutes, I had to poop.
I went to go do that, and my body produced one of the healthiest, longest, most amazing Bristol 4s I had ever seen. It looked like a damn garden hose. As I was admiring it before I flushed, that was the exact moment when the music-while-you-wait ended, and the voice came through on the other end of the line: Thank you for calling Southwest Airlines.
No probability calculator can put together that scene of life. The beautiful Bristol 4. The gut health that I healed with the help of ChatGPT and the yogurt I’ve been forcing myself to eat for two weeks. The layer of water that always ends up on top of the yogurt, which makes the whole process appalling.
AI doesn’t notice the water, or the Bristol 4, or how the person on hold will always answer right when you need to flush the toilet and wash your hands.
It’s improbable.
Speaking of yogurt…
I’ve been contemplating community culture building lately, because a challenge I have with Humans 2.0 is to inject more solar energy into a space that doesn’t have a shared objective or purpose.
This disgusting yogurt has, for whatever reason, inspired me about the word “culture.”
Yesterday I wrote on LinkedIn:
“Community building is like fermentation. I think about making yogurt or kombucha.
I think about the word "culture," and how "culture" for humans and yogurt are pretty much the same thing. If you add one microorganism to milk, it may not establish a culture. If you add yogurt to it, which contains large populations of bacteria, you’re very likely to get more yogurt.
Think about an organization... Building a new culture doesn't work if you only bring in one new human. You need at least three, five, seven.
We can learn a lot from simple things.”
If I were to ask AI to write about culture building in community, I doubt it would reach for Oikos Pro. The same way I talked about olive oil, and sauerkraut to represent various aspects of consciousness.
The generation does not account for that improbability.
And notice, I didn’t tell you what happened with Southwest, because who cares? The point was the image of my eyes admiring the giant Bristol 4 right as the woman answered the phone to reschedule my flight… Which, as a human, you contextually know is already a giant pain in the ass. My telling you all flights out of NYC airports were eliminated was a data point where I trusted you. I left shit out. You felt it anyway.
Life is highly chaotic, sometimes neurotic, often erotic.
Creativity is birthed from loins.
The improbability is what makes a piece of writing amazing.
That’s also what makes life amazing.
You’re asking a very expensive predictability calculator to predict something, and then you’re wondering why you’re feeling bored.
Stay beautiful.