Why Autocorrect Learned The Wrong Word

The keyboard has decided that a particular misspelling is what you meant, and now defends it. The mechanism that adapts to you is the same one that has gone wrong.

Correction is prediction, not spelling

A modern keyboard does not check words against a dictionary and flag failures. It estimates the most probable intended word given the keys pressed and the words around them.

Two models feed that estimate. One describes which letters are likely to be hit by mistake given the physical layout, and one describes which words are likely to follow which.

Because it is probabilistic, it will confidently produce a real word that you did not type, which a spell checker would never do.

Your typing becomes training data

Keyboards adapt to the individual, building a personal vocabulary from what you type so that names, slang and technical terms stop being corrected away.

That adaptation cannot tell an intentional coinage from a typo you did not notice. Both arrive as text you wrote and were apparently satisfied with.

Type a mistake a few times without correcting it and it earns a place in the personal dictionary, after which it competes with the correct spelling.

Overriding once is recorded as a preference

Rejecting a suggestion is a strong signal, because it takes deliberate effort. Systems weight it heavily, which is exactly what you want most of the time.

The trouble is that people also reject suggestions accidentally, by tapping the wrong part of the bar or continuing to type through a prompt.

Those accidents are recorded with the same confidence as deliberate choices, and a small number of them can shift the model's preference permanently.

Undoing a correction sends a muddled signal

When you fix a word the keyboard changed, the system sees a correction being reversed, which it may read as evidence about the word or about the context.

Different implementations interpret that differently, and few of them treat it as a clear instruction never to make that substitution again.

The absence of a plain mechanism for saying no is the gap users feel most, and it exists because the interface has room for suggestions and not for rules.

Names and jargon break the model

Language models are built from general text, where ordinary words are far more common than surnames, place names and specialist terms.

A name that resembles a common word will be replaced by it repeatedly, because the model is behaving correctly according to everything it has learned.

The only fix is to teach the exception, which means adding it to the personal dictionary explicitly rather than hoping repetition will eventually win the argument.