How next-word prediction actually works on iPhone
The suggestion bar sometimes guesses the exact word you were about to type before you have typed a single letter of it. What it is actually built from, why it goes blank at the start of a message, and why the same feature can look eerily sharp one moment and completely lost the next.
Finish a word, and sometimes the box beside it already holds the exact word you were about to type next — not a completion of what you wrote, a guess at what you had not written yet. It can feel like the keyboard is reading ahead, and a moment later the same bar offers something with no obvious connection to the sentence at all. Both of those are the same feature behaving normally. It is not reading meaning any more than autocorrect is; it is doing a narrower, more mechanical job than either result makes it look like.
A different job from autocorrect, sharing the same bar
Autocorrect and next-word prediction sit in the same strip above the keys, which makes it easy to assume they are one feature. They are not. Autocorrect looks backward: it judges the word you just finished typing against a dictionary and swaps it if something close enough and more common exists. Next-word prediction looks forward, at a word you have not started at all, and asks a completely different question — given what is already on the screen, which word tends to come next. One is correcting spelling. The other is guessing content. A word can be spelled perfectly and still get a hopeless next-word guess sitting beside it, because the two boxes were never doing the same kind of work.
What the guess is actually built from
Underneath, prediction is not reasoning about your sentence — it is leaning on which words are commonly typed after other words. “Thank” is very often followed by “you.” “Let me” is very often followed by “know.” A keyboard that has effectively seen an enormous amount of ordinary typing has a strong sense of which word usually comes next after a given one or two words, and it offers whichever candidate that pattern points to hardest. There is no understanding of your particular sentence in that judgement, only a bet on what people typing something similar tended to write next.
Why the bar has nothing to say at the start of a message
Open a blank text field and tap the suggestion bar before typing anything, and it usually has little worth reading, or falls back to the same handful of generic openers regardless of who you are texting. That is not the feature failing — it is the feature with no material to work from yet. A next-word guess needs a word, or at least the start of one, already sitting before the cursor to have anything to key off. The QuickType bar itself always shows three boxes no matter what, but for the prediction half of that bar specifically, an empty field is close to a blank slate: nothing preceding, nothing to predict from. The guess sharpens the moment there is something behind the cursor to read, which is why it feels far more useful mid-sentence than in the first second after you open the keyboard.
Why it learns your phrasing, not only the dictionary’s
iOS keeps a personal, learned layer of typing behaviour alongside its built-in dictionary — the same layer that lets autocorrect eventually stop mangling a name once you have corrected it enough times, cleared only by the Reset Keyboard Dictionary option and by nothing else. Next-word prediction draws on that same personal history, not just a generic baseline everyone’s phone ships with. If you close messages the same way every time, or your particular job or hobby means certain word pairs come up constantly for you and almost nobody else, prediction on your phone gradually leans toward those pairings specifically. That is why the same keyboard can feel noticeably sharper on a phone that has seen months of your actual typing than it did on the day it was set up.
Why it sometimes reads your mind and sometimes reads nothing at all
Both experiences come from the same mechanism, just fed a different strength of signal. A common phrase — “see you,” “talk to you,” “thank you so” — has an overwhelmingly dominant next word behind it, so the guess lands with real confidence and looks like mind-reading. An unusual turn of phrase, a sentence that trails into something nobody types that way very often, or the very first word of a message gives the same mechanism almost nothing strong to lean on, so it falls back to whatever is generically common and the suggestion reads as random. It is not getting worse in that moment. It is running out of signal, and offering its best guess anyway rather than an empty box.
It is still only looking at what is already typed
Next-word prediction has no view of where a sentence is headed, what the rest of the message is about, or what you plan to write after the word it is currently guessing. It works from the words immediately behind the cursor and nothing further — not the paragraph above, not the conversation’s subject, not your intention. That is the same boundary autocorrect runs into from the other direction: a keyboard can measure how words are typically used together without ever knowing what a particular sentence is actually trying to say, and both features stay inside that boundary on purpose, because reasoning about meaning is a different, much heavier job than either one is built to do.
Where Kibo fits
Kibo runs its own next-word prediction from its own settings panel, separate from its Damerau-Levenshtein autocorrect, so the bar still has something to offer right after you press space rather than only when a word is being corrected. Switching between Kibo’s 30 typing languages switches its whole typing-help stack along with it, the same way moving between separate iOS keyboards does, so a prediction built from English phrasing is not the one offered while you are typing in Spanish or Japanese. None of it needs Full Access — showing a suggestion is drawing text on the keyboard’s own toolbar, not reading the clipboard — and Kibo makes no network requests, so whatever it learns from your typing stays on the device. It is free to download.
Frequently asked questions
Is next-word prediction the same feature as autocorrect?
No. Autocorrect judges the word you just finished typing against a dictionary and swaps it if a closer, more common match exists. Next-word prediction guesses a word you have not started yet, based on which words commonly follow the ones already on screen. They sit in the same suggestion bar but do different jobs.
Why does predictive text sometimes suggest a word with no connection to my sentence?
It happens when there is little pattern to lean on — an unusual phrase, or the very start of a message with nothing behind the cursor yet. The keyboard still fills the suggestion slot, so it offers whatever is generically common rather than leaving the box empty.
Does next-word prediction get better the more I use my phone?
Yes. It draws on the same personal, learned layer of typing behaviour that lets autocorrect eventually stop fighting a name you type often, so word pairs that come up a lot in your own typing get weighted toward you specifically over time.
Why is the suggestion bar empty or generic right when I open a new message?
Next-word prediction needs a word already sitting before the cursor to have anything to guess from. At the very start of a blank field there is nothing preceding yet, so the bar falls back to generic openers until there is something typed for it to work from.
Try it on your own keyboard
Kibo is a free custom keyboard for iPhone with real mechanical key sounds, a wardrobe of themes and RGB backlight effects. No account, no ads, and it works with Full Access left off.