Describe, don’t dictate
An autocorrect decides what counts as correct. This keyboard starts from the other end: what matters is what you actually write. That idea has a name, it is older than phones, and we did not invent it.
Two trades, and this one chose the second
There are two ways to deal with a language. The first says how people ought to write: the work of the grammar book, the copy editor, the style guide. It is called prescriptive.
The second observes how people actually write, without correcting them: the abbreviations, the dropped punctuation, the words invented along the way. It is called descriptive, and it is what linguists do.
Both are useful, and neither is a camp you have to join. But a keyboard is not a copy editor: it isn’t proofreading your manuscript, it is with you in a conversation. Correcting someone while they speak goes by another name.
So this keyboard is descriptive. It doesn’t claim to know what is correct. It learns what you write, and helps you write it faster.
Your way of writing has a name: your idiolect
Linguists call idiolect the use one single person makes of their language: vocabulary, turns of phrase, abbreviations, punctuation. Two people in the same family don’t share one.
It is how this app’s App Store page opens: we don’t all make the same typos. A statistical autocorrect aims at the middle of everyone — and the middle of everyone is nobody’s idiolect. That is why it misses you, specifically, so often.
A rule you write aims at you alone. It cannot get the wrong person.
What that word does not cover
Worth being straight about what the term means: an idiolect is a matter of language, not of fingers. It says nothing about the geometry of your thumb.
Yet that is the other half of the job. This keyboard also remembers the exact spot where your finger lands on each key — which is not the centre, and is not the same as mine.
Two layers, then, and it pays not to confuse them: what you write, which belongs to language, and how you type it, which belongs to movement. The app works on both. It does not pretend they are the same thing.
How a decision gets made here
A design is defended with numbers, or it isn’t defended at all. Three habits, and you can check every one of them on this site:
- Look for the measurement before deciding, not after.
- Give the source, so you can check without having to take our word for it.
- Publish what contradicts us. The “What we’ve read” page carries a finding that argues against this keyboard: autocorrect does help typing speed. It is there because it is true.
And when a number is a judgment rather than a measurement, it is labelled as one: the language map says outright that nobody, anywhere, measures which language the world texts in.
The studies we’ve read, with their sources →
Our own numbers, with their limits →
No corpus — and that is something given up
To study a language, you gather a corpus: large amounts of real text, taken from real people. It is the raw material of the trade — and a keyboard app would be perfectly placed to build one.
This one gathers none. Nothing you type ever leaves your phone: not a word, not a sentence, ever. There is no server, on the developer’s side, for your text to pass through.
It costs something, and it should be said: we will never know how “people” in general write. What the keyboard learns, it learns on your phone, for you, and it stays there.
Even the anonymous measurement the app may ask you to accept — off by default, and revocable whenever you like — carries fourteen counters and nothing else: not a word, not a key, not a typing rhythm. Nobody could build a study of language out of it. Least of all us.