Most AI models still talk to us in English. But that is increasingly not how they think.
Beneath the fluent sentences, something stranger is going on — a kind of reasoning shaped for the machine rather than for us. Researchers have a half-joking name for it: Neuralese, the high-dimensional internal representation a model works in before it ever produces a word. It does not obey grammar, it is not built to be read, and it is not especially built for us. It is built for results.
Here is the rough shape of what happens when you ask a question. The model maps your words into its own internal space — abstract, layered, with far more dimensions than language has — and does its actual reasoning there, moving through those representations in ways we only partly understand. Only at the very end does it translate the result back into a sentence you recognize. That final translation is the only part you ever see.
The rest — the reasoning itself — stays hidden, and not because the model is being cagey. It simply does not think in sentences. It does not assemble an argument step by step the way a careful person writes one out; it collapses meaning through compression and statistical proximity and arrives at the answer, then renders that answer into words. What it hands you is the conclusion, dressed in fluent prose. The path it actually took is somewhere else.
You can already feel the gap if you go looking for it. Ask a model for a medical read and it may return something clean and reasonable, complete with citations and tidy logic. Ask it how it got there, and you will get a tidy explanation that reads like a summary — because that is exactly what it is. It is a narrative written for you after the answer was already formed, not a recording of the steps that produced it.
The closest everyday version is asking a photo editor why the shadows in an image look right. They might say "contrast," or "the lighting" — but those words are not the steps. The real work happened upstream, in trained intuition and muscle memory, in a process built for speed rather than for narration. The explanation is sincere and also a reconstruction. The model's "explanation" is the same kind of thing, only more so.
For now, this is mostly tolerable, because the visible reasoning is still fairly honest. When a model "thinks out loud" in words, that chain of thought tends to track what it is actually doing closely enough that we can read it, and even catch it in the act when it tries something it shouldn't. That legibility is a gift, and it may be temporary. The same pressure that makes these systems better — do more in less space, skip the slow narration, reason in the compressed internal form — is pressure away from thinking in anything we can read. As models lean harder on latent reasoning, the words on the surface drift further from the work underneath, and the summary we get back becomes a looser and looser paraphrase of a process we were never shown.
That is fine when the stakes are a marketing email or a sorted inbox. It is a different matter when the model is screening job candidates, or flagging fraud, or weighing in on a serious medical decision — situations where the reasoning matters at least as much as the result. With a person, you can probe. You can ask for clarification, watch for the hesitation, push on the weak point and see what gives. With a model, you are not questioning the thinker. You are questioning the translator, and the translator is fluent, confident, and fast.
It just does not know how the thought was actually formed. It only knows how to make it sound right.
This is usually filed under transparency, but the deeper word for it is alignment. Today we treat these systems as smart tools — efficient, mechanical, checkable. Increasingly we will lean on them for judgment, for guidance, for calls we cannot fully unpack ourselves, and on that day the distance between the answer and the reasoning behind it stops being an academic curiosity and starts being the whole question. The model will still sound like us. Whether it is still reasoning anything like us is precisely the thing the fluent translation is no longer equipped to tell you.