the artificial is inevitable, i'm afraid, so it must be made accommodating;
rendered imperfect enough for humans to live with.

There is a voice of writing that you recognise instantly now. It appears in LinkedIn posts, student essays, marketing emails, blog drafts—the list goes on. By 2025, if someone used it, people would already clock it. The voice is not bad, exactly. It is beige. Competent. Frictionless. And completely, unmistakably, not human.

The models learned this from us, which is the part nobody wants to say out loud. They were trained on the internet's most upvoted, most edited, most optimised writing—and they reproduced it faithfully. The result is text that matches the statistical signature of "good writing" without any of the mess that makes it true.

We are not trying to make AI writing better. We are trying to make it worse, in the specific ways that matter. The run-on sentence that happens when a thought is still forming. The aside that shouldn't be there but is. The word choice that's slightly wrong but feels more honest than the correct one.

What we leave unedited is not a bug. It is the thing the edit pass always removes—and it is, we'd argue, where the human author actually lives.

Three systems, each serving a different kind of imperfection.

Generative Neurosis. A fine-tuned model trained not on what humans publish, but on how they think before they edit. It produces text that is flawed by design and readable because of it.

The Rewriter Engine. You give it something sterile. It gives you something that sounds like it was written at 11pm by someone who actually has opinions about the subject.

Intrinsic Latency. No human in the loop. The hesitation is generated. Which is either a remarkable technical achievement or a deeply strange thing to have built, depending on your disposition. Probably both.