What Is Word Prediction? Definition and How It Works

Word prediction is a feature that suggests the next word or phrase as you type, before you have finished your sentence. It works by calculating the most probable next word given the preceding context - a language model probability problem. Used across smartphones, keyboards, and writing tools, word prediction reduces keystrokes and helps writers maintain flow. Pressing Tab in Charm, or tapping a suggestion on a phone, accepts the prediction.

How does word prediction work technically?

Word prediction is fundamentally a next-token prediction task. Given a sequence of words already typed, the model assigns a probability to every possible next word in its vocabulary and surfaces the highest-scoring candidate.

The simplest implementation uses an n-gram model. A bigram model looks at the single preceding word; a trigram model looks at the two preceding words. Given "the quick brown", a trigram model calculates the probability of "fox", "dog", "cat", and every other word in its vocabulary following that exact three-word sequence, based on how often each pairing appeared in the training corpus. N-gram models are fast and require minimal compute, but their short context window limits accuracy on longer, more complex sentences.

Modern systems use transformer-based language models - the same architecture behind large language models like GPT. Transformers process the entire preceding context, not just a fixed n-gram window. This allows predictions to be informed by meaning established many sentences earlier, producing suggestions that feel more natural and contextually aware.

The practical output is determined by two thresholds: context length (how many preceding characters or words the model reads) and confidence threshold (the minimum probability score required before a suggestion is shown). Charm's Oracle reads up to 100 characters of prior context and only surfaces a suggestion when confidence is high enough to be useful - preventing the constant interruption of low-quality guesses.

What is the difference between word prediction and autocomplete?

The distinction matters, and the two terms are often conflated.

Autocomplete finishes a word you have already started. If you type "bec", autocomplete might suggest "because" or "become". It is a within-word operation, matching your partial input against a dictionary of whole words.

Word prediction operates at the space between words. When you have just finished typing a word and pressed space, word prediction suggests what comes next - before you have typed a single character of the new word. It is a between-word operation, based entirely on context rather than partial input.

Both can appear in the same interface. Smartphone keyboards, for example, typically show three suggestions above the keyboard that shift between autocomplete (if you have started a word) and word prediction (if you have just finished one). Research shows smartphone users accept word prediction suggestions approximately 40% of the time, making it one of the highest-impact typing assistance features available.

How is word prediction used in writing tools?

Word prediction finds its strongest use cases in three scenarios: high-volume repetitive writing, language learning, and accessibility.

For repetitive professional writing - support emails, legal documents, code comments - word prediction shortcuts recurring phrases. Studies in professional writing contexts show that word prediction reduces keystrokes by 15-25%, translating to meaningful time savings over a full working day.

For non-native English speakers, word prediction assists with collocations: the conventional pairings of words ("make a decision" rather than "do a decision") that are difficult to internalise from grammar rules alone. Seeing the predicted word validates or corrects an instinctive choice in context.

For users with motor difficulties - RSI, Parkinson's, or limited fine motor control - word prediction dramatically reduces the number of key presses required. It is a core accessibility feature in augmentative and alternative communication (AAC) devices, where minimising input effort is critical.

Charm's Oracle feature brings system-wide word prediction to Mac. Because Oracle uses the Accessibility API rather than any app-specific integration, it works in every text field - including Slack, VS Code, and Discord, where no native Mac prediction exists. When Oracle generates a suggestion, the predicted word appears with a purple glow inline. Press Tab to accept it; keep typing to dismiss it.

Is word prediction private?

Privacy varies significantly by implementation, and it is worth understanding the difference before choosing a tool.

Cloud-based prediction systems generate suggestions on remote servers. Your text is transmitted, processed, and returned with suggestions in milliseconds. The round-trip is fast enough to feel instant, but your writing - including draft emails, private notes, and anything you type into a password field or confidential document - has left your device.

On-device prediction keeps the language model on your Mac. No text is transmitted. The model runs entirely within your hardware, producing suggestions without any network request. For writers handling confidential information, this distinction is significant. Charm's Oracle is fully on-device by default. Even the optional OpenAI API key (used for Charm's enhanced grammar feature, Polish) does not affect Oracle - word prediction remains local regardless of any cloud settings.

Frequently asked questions

What is word prediction?

Word prediction is a feature that suggests the next word before you begin typing it, based on the language model's calculation of the most probable continuation given your preceding text. Accepting a suggestion typically requires pressing Tab or tapping the displayed word.

Is word prediction the same as autocomplete?

No. Autocomplete finishes a word you have already started typing. Word prediction suggests the next word before you have typed any of it. Both can appear in the same interface, but they operate on different triggers and use different mechanisms.

Does word prediction improve over time?

Cloud-based systems can improve by incorporating usage data from many users. Charm's Oracle does not update its core model from your writing, but learns your accepted patterns locally - without sending any text to external servers.

Is word prediction private?

Only if it runs on-device. Cloud-based tools send your text to remote servers to generate suggestions. Charm's Oracle runs entirely on your Mac - no text leaves your device, even for confidential documents or private notes.

How do I enable word prediction on Mac?

macOS does not include system-wide word prediction. Install Charm and grant Accessibility access to enable Oracle. When Oracle has a suggestion, the predicted word appears with a purple glow inline. Press Tab to accept it, or keep typing to ignore it.