Careful. Well-organized. Nothing out of place. And yet, nothing sticks.

I wrote before about why AI-generated text feels off. This time, I want to flip it around.

Why does some writing feel like it could only have come from that person — and earn your trust?

Starting from "is it a problem if we can't tell it apart from an LLM?" and going deeper to "what actually makes writing feel like someone?", here's what I think does it.

A Distinctive Voice

LLMs output "the most statistically likely next word given context." That's why they converge toward the average. They write in a voice that sounds like everyone — and no one.

Someone who has written for years develops their own fingerprint. The way they break sentences. The conjunctions they reach for. The dashes they use, or don't.

These form unconsciously. A few lines in, you know exactly who it is.

That voice can't be reproduced from a short prompt. It takes thousands of pieces, hundreds of posts, accumulated over time.

(For the record: I have no idea what my own writing fingerprint actually is.)

A Punchline That Lands

Am I alone in finding AI writing unnecessarily long? There are no creative punchlines — no moments that knock your head off. That's structurally predictable from how LLMs work.

But if a piece of writing has a line that hits hard enough to make you lose your head, you trust unconditionally that a human wrote it.

Specific Experience

LLMs are trained on the collective intelligence of the internet. They're great at things everyone has written about. What they don't have is your life.

"When we recorded last Tuesday, Hikaru said this to me—" earns trust not just because it's specific. It's because the choice of episode reveals character. An LLM doesn't have access to your memory — or your sense of what matters enough to share.

The flip side: words disconnected from lived experience, which LLMs tend to produce a lot of, always feel slightly weightless.

Specific Relationships

Relationships are closed contexts. Who you choose to write about is one of the most revealing things about you.

In my piece on AI and dividual identity, I referenced Keiichiro Hirano's idea that the bunjin (分人, "dividual" — the partial self that emerges in each relationship) is born from relationships. AI is a mirror — frictionless. No real encounter happens. So interacting with AI doesn't grow the kind of self that human relationships do.

(I actually read Hirano's book, by the way.)

Discontinuity Over Time

LLMs have no timeline. They give you "the best answer for right now," with no contradictions from the past, no record of change.

Human writing has change baked into it -- and occasional contradictions too. "I used to think X, but now I think Y" is a record of living through time. That kind of reversal, spreading a little embarrassment around as you go, is part of what makes someone feel human.

Imperfection and Uncertainty

LLMs try to produce complete answers. Ambiguity, hedging, unresolved tension — they're not great at leaving those in.

"I'm not sure, but this is what I feel" is actually the most honest reflection of how humans think. Only someone who respects their own thinking can present an unresolved question as is.

Sometimes rougher, more uncertain writing feels more trustworthy than something polished to a high shine. Because it looks like the traces of someone actually thinking.

Trust Signals Stack

None of the elements above are individually enough to make writing feel human. One alone doesn't do it.

But they compound. The more they overlap, the stronger the signal of personality.

I've been trying to capture more everyday moments on this blog lately, and maybe — just maybe — that's an intuition about building trust through the accumulation of posts that each carry something of myself.