Tech

How autocorrect learns the way you type daily

Rive News Research Desk

Published October 28, 2026 · 9 min read

By Rive News / Research desk | Rive News / Published October 28, 2026 at 9:00 AM PKT

Key takeaways

Autocorrect combines a general language model, your personal dictionary, and a keyboard model that scores typos by key distance.
Every accepted suggestion, manual correction, and added word teaches it, which is why it can also learn your mistakes.
You can retrain it by adding words to the personal dictionary, removing bad suggestions, or resetting its learned data.

[A woman texts on a smartphone.] Photograph: the subject

Image Credits: Wikimedia Commons

Autocorrect feels personal because, over time, it becomes personal. A new phone corrects you toward standard English. The same phone a year later knows your friends' names, your slang, and the technical terms you type at work.

That change does not come from one clever trick. It comes from three layers working together: a general model of the language, a personal record of your words, and a model of your keyboard itself. This article explains each layer, how they learn from you, and why they sometimes learn the wrong thing.

The general language model

The first layer is the language model, a statistical map of which words tend to follow which other words. It is built from large collections of ordinary text, and it answers questions like: after the words "I will meet you at the", is "office" more likely than "quantum lab"? Almost always, yes.

A common version of this model is the n-gram. An n-gram looks at the last two or three words to predict the next one.

That short window is powerful for common phrases, but it is also the reason autocorrect can fail badly on a single word with no context. With nothing before it, the model falls back on raw word frequency, and a common word beats a rare one even when the rare one is what you meant.

This layer also explains why autocorrect seems to prefer standard language. It was trained mostly on formal writing, which contains far less slang, dialect, and technical jargon than your messages do. When you type a word the model has rarely seen, it assumes a typo for a more common word and offers that instead.

The general model is updated by software updates and, on some keyboards, by federated learning. Google's Gboard, for example, uses federated learning to learn new words and phrases: the phone learns locally while it charges, connects to Wi-Fi, and sits unused, then sends what it learned to Google, where it is combined with learnings from other users.

According to Google, the text you type is not sent in this process, only the learned patterns. Federated learning is on by default in Gboard, but you can switch it off.

The personal dictionary

The second layer is the one that makes autocorrect yours: the personal dictionary, sometimes called the learned words list. This is a record of words you have typed, accepted, or added yourself, stored on the device.

It learns in several ways. When you type a word and tap it in the suggestion bar, you confirm it. When you correct a suggestion back to your word, you teach the model that your word wins for that input.

When you tap a word underlined in red and choose Add to dictionary, as Google's Gboard help pages describe, you add it explicitly. Over time, those signals shift the odds so that your words outrank the general model's defaults.

That is how a friend's unusual name stops being corrected. To the general model, a name like Aislynn is a highly improbable letter sequence, and Ashlyn or Aileen look far more likely. Once the name is in your personal dictionary, the model has your data to outweigh the general statistics, and the corrections stop.

The same mechanism explains how autocorrect learns mistakes. If you tap a wrong suggestion by accident, even once, the system can log that preference and weight it in future predictions. Repeatedly correcting a word the wrong way teaches the wrong pattern just as effectively as correcting it the right way teaches the right one.

Autocorrect does not know what you meant. It only knows what you accepted.

The keyboard model

The third layer has nothing to do with language. It is a model of your keyboard's physical layout, and it scores typos by how far your fingers slipped.

On a QWERTY keyboard, the letters o, i, and p sit next to each other. A typo like "wprk" is far more likely to be corrected to "work" than to "park", because the mistyped letter is adjacent to the intended one. The classic transposition "teh" is corrected to "the" so reliably because it is an adjacent-key swap the model is built to recognise.

This layer also explains why some typos are never corrected. If your slip lands on a real word, the language model may accept it happily.

Typing "duck" quickly with a finger slip can produce a different real word, and if that word is also more frequent in your own messages, the personal layer may back it up. The result is a correction that looks absurd but follows perfectly from the model's logic: key distance plus frequency plus your own history.

How the three layers decide together

When you type a word, the keyboard runs all three layers at once and combines their scores. The language model proposes likely words for the context.

The personal dictionary boosts words you have used before. The keyboard model scores each candidate by how plausible the typo is, given the keys you actually pressed.

If the input matches a word in your dictionary, it usually survives untouched. If it matches nothing, the system offers the highest-scoring candidate, and may apply it automatically when you press space or punctuation. If several candidates score closely, you get a suggestion bar instead of an automatic change, so you can pick.

Context narrows the field further. Some keyboards look at the last five to ten words, not just two or three, to predict the next one.

Type "The API returned a 404" and "error" will outrank "status" or "message", because the surrounding technical context shifts the odds. Type the same fragment in a cooking chat and the ranking changes.

This combination is also why autocorrect struggles with code-switching. If your keyboard is set to English, it will try to interpret every word in another language as a mangled English typo, because the English language model has no patterns for those words. Multilingual keyboards and per-language dictionaries help, but only if the keyboard knows which language you are typing in.

Young people typing on smartphones using their thumbs

[Young people type on smartphones using their thumbs, producing the characteristic slips the keyboard model scores.] Photograph: the subject

Image Credits: Wikimedia Commons

Why autocorrect goes wrong, and how to fix it

Most autocorrect failures fall into a few patterns. The first is accidental learning, where one mistaken tap teaches a wrong preference that then repeats.

The second is overcorrection of names, slang, and technical terms that the general model has rarely seen. The third is regional mismatch, where a British spelling is corrected to an American one, or vice versa, because the keyboard's default language does not match yours.

Fixing the first pattern starts with the suggestion bar. On Gboard, touching and holding a suggestion you do not like and dragging it to the trash removes it. On iPhone, tapping a suggestion and correcting it back teaches the opposite preference, though it may take several repetitions to outweigh a strongly learned mistake.

Fixing the second pattern means adding words deliberately. Open the personal dictionary in your keyboard settings, choose your language, and add the names, terms, and abbreviations you use.

Gboard's help pages describe adding a word by typing it, tapping it when underlined in red, and choosing Add to dictionary. That single action does more than weeks of fighting individual corrections.

If the whole system has gone off track, resetting is the cleanest option. On iPhone, resetting the keyboard dictionary clears learned words and restores the defaults.

On Android, Gboard's privacy settings let you delete learned words and data without removing the keyboard itself. After a reset, expect a short period where autocorrect feels generic again while it relearns your vocabulary.

Privacy is part of the same trade. Gboard's personalisation stores what it learns on the device, and you can delete that learned data in its privacy settings.

Federated learning sends only learned patterns, not your typed text, according to Google, and it runs only while the phone charges on Wi-Fi and sits unused. If you would rather not contribute at all, switching federated learning off leaves the on-device learning intact.

The permission side of keyboards deserves the same attention as any other app. A keyboard sees everything you type, which is why its privacy settings matter more than most apps' settings. For a broader look at why apps request the access they do, see why apps ask for so many permissions.

Frequently asked questions

Does autocorrect actually understand what I mean?

No. It calculates probabilities from word frequency, your personal dictionary, and keyboard layout. It does not understand meaning, which is why it can confidently replace a correct word with a more common one.

Why does autocorrect keep changing a name I type often?

Because the general language model has rarely seen that name, so it scores a common similar word higher. Adding the name to your personal dictionary shifts the odds in your favour.

Can I teach autocorrect a new word permanently?

Yes. Add it to the personal dictionary in your keyboard settings, or tap it when it appears underlined and choose Add to dictionary. From then on, the keyboard treats it as a known word.

Why does autocorrect learn my typos?

Because it learns from acceptance, not from intent. If you tap a wrong suggestion, or repeatedly type a misspelling without correcting it, the system records that pattern as your preference.

Does turning off autocorrect delete what it learned?

Not necessarily. Turning the feature off stops corrections, but learned words may remain until you reset the keyboard dictionary or delete learned data in the keyboard's privacy settings.

The bottom line

Autocorrect learns the way you type through three layers: a general language model that knows how words usually follow each other, a personal dictionary that records the words you actually use, and a keyboard model that scores typos by how far your fingers slipped. Every accepted suggestion and manual correction adjusts those odds, for better and for worse.

The keyboard is not stubborn. It is simply faithful to the pattern you taught it.

Tags: autocorrect, keyboards, typing

Rive News is a research desk covering everyday technology. It writes explainers from public sources and documents, not from first-hand product testing.

Subscribe to the Rive News newsletter for one plain-English technology explainer each week.

Rive News follows a corrections policy: errors are corrected openly and the article is updated with a timestamp.

Taggedautocorrectkeyboardstyping