A friend told me recently that LLM writing feels dead to him, while human writing feels alive. I think it’s because an LLM doesn’t learn or live in time and in an environment, the way you and I do. 

You remember, at least roughly, when you learned something. Stuff you picked up as a kid, different place, different people… it has its own vibe, its own web of associations. It’s totally unlike whatever you learned last week at work. 

An LLM has no such equivalent. It does not learn in time but all at once. It can tell you which event came first, but it doesn’t experience one as a memory and the other as a recent discovery. 

And this claim is easy to vet from experience as well, so much so that it’s a running joke on X at this point about asking an LLM for a time estimate. It’ll say two weeks, and then you’ll have it do the thing in two hours. That happens because its estimates come from training data about how long things take humans. 

It has never done anything and noticed how long it took. 

It has never finished something difficult and noticed that the afternoon is gone. 

It doesn’t know what “two weeks” feels like.

But a human, writing, does. It happens while someone is waiting, losing interest, deciding whether to trust you — even when that someone is a person you’re talking to inside your head. It happens in a real place, between actual people, in whatever mood is actually operating at that moment. And I think writing begins to feel alive when you can sense the writer adjusting to those particulars.

We all do this naturally! Say you’re the jester of your friend group. You crack some jokes, but then you get serious. You don’t perform all night. You say something funny, then play it cool, maybe shut up for a while and just read the room, then say something non-ironic, and only then land another joke. Every actually funny person gets this. In fact, the funniest people are especially good at knowing when to stop being funny. Even comedians, in settings where everyone has explicitly come to laugh, interrupt a comic bit with a moment of sincerity so their audience can catch a glimpse of the human under the performance. Their sensitivity to the room produces variation.

Some friends can hear me work out an idea in a long ass voice note where I’ve left all the uncertainty, fillers and pauses in. But with someone I feel I have to impress, I work the thought out alone and send the cleaner version. Often, I just type it out. So, even the texture of the communication changes with the relationship.

I think HRV is a decent analogy here. Heart rate variability is a marker of health, and all it measures is how inconsistent your heartbeat is at the micro level. A healthy heart is not a metronome. Its beat changes because the body is continually giving it new information and it keeps adapting to what’s happening around it.

But even irregularity by itself can be regular, like noise is irregular. However, what I’m pointing at is the kind of variation that can only be produced by feedback from a real, living situation. Human expression is shaped not just by what the writer senses, but by what saying the thing might cost them. They may expose themselves, damage a friendship, look foolish, change their own mind, or discover halfway through that they don’t believe what they intended to say. That vulnerability forces creative regulation. The words cost the person producing them. 

But an LLM has no sense of risk or cost. Ask it to write anything and it stays perfectly consistent the whole way through and overindexes on whatever you asked for. No human writes like that. Ask me to be funny and I won’t be funny all the way through, because it gets tiring, and good authors know the dosage and the timing. An LLM doesn’t, because it has no idea what tiring means.

I’m guessing the root of the problem begins lower down, inside the mechanism itself. An LLM is a next-token predictor. Each choice alters the field of likely choices that follows. Once the model establishes a cadence, a register, or a shape for the sentence, its own prose begins pulling it further in that direction. The prose falls into a groove, and then the groove starts writing the prose. The writing settles into a locally stable pattern, and everything it has already written becomes another reason to remain there.

But Mustafa, human writers fall into grooves, too? Yes, and much dead human writing comes from exactly that. However, a human author can experience the groove as a groove!

A sentence I’ve just written returns to me as something I’m newly perceiving again myself. I can hear the last three sentences in my head and figure out my writing is getting too monotonous. I can notice when I sound preachy, unserious, overbearing, verbose. And then I can change it to something more desirable. Heck, sometimes I will delete many a good sentence because it has started overshadowing the rest of the paragraph.

The point is, I don’t just squeeze out words like toothpaste. I react to them as both author and audience in realtime.

Sure, with agentic loops, an LLM can be instructed to imitate this. I can ask it to do multiple passes to vary the rhythm or insert an aside. But even this manufactured variation settles into its own tedious little loop. The model can imitate surprise and irregularity to the point of the prose being quite unsurprising and regular in its irregularity. It can perform responsiveness in a way that is itself curiously unresponsive.

Because there’s no real feedback from an environment, in time! It’s just trying to simulate it. What you actually need is some kind of independent pressure, something outside the text itself that’s able to say, look, this pattern has done what it came here to do, and what it came here to do is now done. An LLM ordinarily has no such counterforce. Its previous language is primarily evidence about what language should come next. 

It asks “What continuation fits what is already here?” A human writer often asks, “Now that this is here, what has changed? And what needs to follow?”

Writing or speech that feels alive is a mind constantly adjusting to the room, to the reader, and to the larger context the conversation inhabits. It comes from a mind sitting in time, inside an environment, sensitive to both. The LLM has neither, so it holds one register at full pressure from the first word to the last. And nobody alive sounds like that!