The Keyboard That Knows Where It Is

You're typing a work email and the suggestion bar offers "pursuant." You switch to Instagram to caption a photo and, within seconds, it's nudging you toward "vibes" and a string of sunglasses emoji. Same fingers. Same phone. Same keyboard app.

Completely different suggestions.

This isn't a glitch. It's one of the more quietly sophisticated things your phone does, and the mechanism behind it is genuinely interesting rather than just vaguely impressive.

Two Different Memory Systems Working at Once

Most people assume autocomplete is one thing: a giant word list, maybe updated occasionally, predicting what comes next. That's roughly half the picture.

Modern mobile keyboards like Gboard, SwiftKey, and Apple's QuickType run two memory systems in parallel. The first is a global language model, baked in at the factory level, trained on billions of words scraped from books, websites, and anonymised text data. It knows that common conversational sequences are far more probable than garbled alternatives. That model doesn't change when you type. It's the baseline, the skeleton.

The second system is personal and contextual. Your keyboard tracks which app is in the foreground when you type and builds a separate statistical profile for each one. Every time you accept a suggestion in Gmail, that acceptance gets logged against a Gmail-specific weight. Every time you tap away from a SwiftKey suggestion in WhatsApp and type something else, that correction updates a WhatsApp-specific profile.

Over time, those per-app profiles diverge. Significantly.

The Mechanic, Worked Through

Picture two people, Mara and Joel, who both bought the same phone on the same day. Eighteen months later, Mara uses hers almost exclusively for work: long emails, Slack messages full of product names and industry jargon, the occasional LinkedIn post. Joel uses his almost entirely for group chats, sports commentary, and memes.

By month eighteen, their keyboards have seen roughly the same number of keystrokes. But Mara's Gmail profile has been corrected toward words like "deliverable" and "stakeholder" hundreds of times. Joel's WhatsApp profile has bent toward team abbreviations, casual shorthand, and strings of laughing emoji. When Mara opens WhatsApp for the first time in a week, her suggestions will be noticeably stiffer than Joel's. The global model is identical on both phones. The per-app weighting is not.

If you're seeing weirdly formal suggestions in a casual app, you're watching that exact gap play out in real time.

What the App Actually Tells the Keyboard

This is the part most explanations skip, which is a shame, because it's genuinely interesting.

When an app opens a text field, it sends a metadata packet to the keyboard. Developers call this the input type and input hints. A password field tells the keyboard to suppress suggestions and autocorrect entirely. An email address field suppresses the space bar's automatic period-insertion and nudges toward completions like ".com" and ".org." A phone number field switches the keyboard to numeric mode without you touching anything.

Beyond those hard-coded types, apps can send softer signals too. A messaging input hint nudges the global model toward shorter, more casual sequences. An auto-complete hint tells it to be more aggressive with suggestions. A no-personalisation flag, which some banking and healthcare apps use for privacy reasons, tells the keyboard to ignore learned preferences entirely and fall back to the generic model.

So the suggestions you see are not just a product of what you've typed before. They're partly a product of what the app developer decided to tell the keyboard about the context. Every time you tap a text box, you're getting a three-way negotiation between the global model, your personal history, and the app's own instructions. Think of it less like a dictionary and more like a translator who's been quietly briefed by the room.

Why the Same Word Scores Differently in Different Places

Under the hood, each suggestion carries a probability score. The keyboard ranks candidates and shows you the top three or five, depending on the interface. That score is a weighted average of several inputs:

  • N-gram frequency: how often this word follows the previous two or three words in the training data
  • Personal correction history: how often you have chosen this word in this app after seeing it suggested
  • Recency: words you've typed in the last few days score slightly higher than words you typed six months ago
  • Input type signal: the app's own hint, which can multiply or suppress certain word categories

Change the app, and the personal correction history component changes entirely. The n-gram component stays constant. The input type signal changes. The result is a different final ranking, even for identical preceding words.

Type the same two-word opener in Outlook and the top suggestion might be a formal professional phrase. Type the same opener in a gaming chat app and it might resolve toward something casual or saltier, depending on your history. Same preceding characters. Completely different posterior probability.

What People Consistently Misread About This

The most common misconception is that keyboards are secretly reading your messages for content in real time. Reputable keyboards don't send your actual text to a server to generate suggestions on the fly. The learning happens on-device, through federated learning in the case of Gboard, where the phone updates a local model and only aggregated, anonymised weight adjustments ever leave the device. The paranoia here is mostly misplaced.

The second misconception is that resetting your keyboard's learned data will give you dramatically better suggestions. Sometimes it helps if your habits have genuinely changed. But what most people experience after a reset is a keyboard that feels dumber for two to three weeks while it rebuilds its per-app profiles from scratch. The global model was never the problem. The personal layer is the part that made your suggestions feel sharp, and you just deleted it.

The third misread: people assume the keyboard is malfunctioning when suggestions feel off in a new app. It isn't. It just has zero personal data for that context yet. Give it a few hundred keystrokes and the suggestions will start bending toward your actual patterns. The global model is the scaffold. Your typing is the building.

The Practical Upshot

If you want better suggestions faster in a new app, accept the offered completions even when they're slightly wrong, then correct immediately after. That correction-and-override sequence is the fastest way to train the per-app profile. Ignoring suggestions entirely and typing everything manually is slower to learn from, because the keyboard is looking for the delta between what it offered and what you actually wanted.

If you type in multiple languages across different apps, a common pattern for bilingual users who switch between a family WhatsApp group and an English-language work Slack, enable multilingual typing in your keyboard settings rather than switching language packs manually. Both Gboard and SwiftKey support simultaneous multilingual models, and keeping them active means the per-app profiles accumulate correctly instead of getting scrambled every time you switch.

It's also a useful reminder that your phone is full of systems like this: quiet, incremental, building a statistical picture of your habits one small correction at a time, in ways you'd never notice unless something went wrong. The keyboard is just the most legible example because you interact with it constantly and the output is right there on screen.

Your keyboard isn't lucky when it predicts you well in one app and useless in another. It's the result of months of arithmetic running in a process you never opened. It knows where it is. It just needed time to learn what you do there.