You can tell. That is the whole problem in three words.
Someone sends you a note and it opens with a scene-setting clause about a fast-paced world. It offers three things in a list, because three is the number lists have. It reaches for a contrast — not merely X, but Y — and lands the contrast on nothing. It closes by offering to help further. Nothing in it is wrong. Nothing in it is anyone's. And those em-dashes, I know you observed them earlier in this paragraph and immediately concluded that this is a piece of AI slop. Well, it's not your fault. We're all in the same boat.
This is not a defect in any particular tool. It is what the tools are for. A large model is trained on an enormous aggregate of written content and then tuned to produce sentences that a very wide readership will find acceptable. Acceptability at that scale has a shape: hedged, symmetrical, mildly warm, allergic to the odd word. The safest sentence is the one that offends nobody, and a sentence that belongs to nobody is the sentence that offends nobody. Averaging is not a side effect of the technology but the technology itself.
For most of what people generate, that is fine. Nobody needs a distinctive voice in a meeting summary. The trouble starts when the same layer of paint gets applied to writing where the person was the point — the cover letter, the condolence note, the investor update, the message to a customer who is already annoyed. In those cases the flatness does not read as neutral. It reads as absence. As much as The tell is this was written by AI, it is this person did not bother.
That is a strange thing to have built. Writing used to be one of the cheapest signals of individuality available: free, portable, hard to fake at length. A decade of homogenised output erodes it in the one direction that cannot be undone by trying harder. If everyone drafts from the same middle, effort stops being legible.
What a writing identity actually is
Linguists have a term of art for this. An idiolect is the speech of a single person: the particular set of habits — vocabulary, sentence length, rhythm, punctuation, the moves someone makes when they are hedging or being direct — that distinguishes one speaker of a language from every other speaker of that language. It is not style in the literary sense. It is closer to a gait.
It is also measurable, within limits. In 1963 Frederick Mosteller and David Wallace(No, not the one from The Office) took the twelve Federalist papers whose authorship was disputed between Hamilton and Madison, out of eighty-five in total, and attributed them statistically using the frequencies of ordinary function words — upon, whilst, by. Not the arguments. The connective tissue. The habits people do not know they have are the ones that identify them.
That study is worth holding in mind for what it required: long documents, two candidate authors, strong priors, decades of scholarly argument afterwards. Stylometry under laboratory conditions is a demanding instrument, and it is a different job from the one we are doing.
The layer we are building
Idiolect AI is the personalisation layer for AI writing. It learns from what you have actually written and brings that understanding to the assistant you already use. When you ask for a draft or a rewrite, it helps the model follow your habits: how quickly you get to the point, how much explanation you give, which words you reach for, how you soften a difficult sentence. You supply the subject and what needs saying. Your writing supplies evidence of how you would say it.
That understanding evolves as you use it. More writing and feedback give Idiolect AI a fuller picture of you, including the differences between how you write to a colleague, a customer and a friend. A first set of examples is a starting point. As the picture develops, there is more to work from than a few broad instructions such as “warm” or “concise”. The aim is less repeated explanation from you and a closer fit to the particular thing you are trying to write.
That understanding can travel between AI tools, so changing assistants does not mean explaining yourself from the beginning. Drafts can also be checked with Voice Match, when there is enough evidence, to see how closely they resemble your writing. It is a relative measure; it cannot establish who wrote something. The practical question is whether the help has preserved enough of your habits for the result to feel familiar. A useful assistant should leave you with less work and something you can still recognise as yours.
id·i·o·lect · the speech of one person.