Is predictive text always based on a large language model?

No. Different systems use different models and methods, including smaller statistical or neural language models.

Must users accept the first suggestion?

No. A suggestion is a candidate; users can choose another or continue typing their own text.

Does prediction accuracy measure typing speed?

No. Speed also depends on reading, selection, correction, device interaction, and the user's task.

AI Glossary / Natural language and speech / Beginner

Predictive text

Definition

Predictive text is a text-entry feature that suggests words or phrases a person may want to type, using information such as entered characters and surrounding language. A user can accept a suggestion instead of entering every character. The feature can complete a partly typed word or propose what might come next.

What it means

A prediction is a candidate continuation, not a statement that the system knows the writer's intention. Several different continuations can fit the same text. After “Please send the,” a writer might mean invoice, address, photo, or something else entirely.

Predictive text overlaps with other keyboard assistance but is not identical to them. Autocorrection changes or proposes a correction to entered text; completion finishes a partial item; next-word prediction suggests a following item. A product may combine these functions, and its interface determines how suggestions are shown, accepted, or ignored.

How it works

The system uses a language model or other ranking mechanism to score candidate words given the available context. Prefix information can restrict candidates when a word is partly typed. Some systems also adapt to a user's vocabulary or patterns, while others operate with more general models.

The interface presents a manageable number of suggestions rather than every possible continuation. Designers decide when to update them, how to display longer phrases, and what action accepts a suggestion. These choices affect usefulness because reading and selecting a candidate also takes attention and time.

Evaluation must therefore distinguish prediction quality from typing benefit. A model may rank the intended word highly, yet a user may enter it faster than they can inspect the suggestions. Longer suggestions can also influence what people choose to write, so testing should consider whether the final message still reflects the writer's intended meaning.

Three practical examples

1. Mobile messaging: A person accepts a suggested everyday word after typing its first few letters, reducing the remaining character entry.

2. Assistive communication: A text-entry system offers candidate words that a user can select through their available input method, with suitability evaluated for that person's interaction needs.

3. Routine workplace writing: A composing tool suggests a recurring phrase, while the writer checks names, dates, and commitments before including it in the message.

Think of it this way

Think of a writing partner quietly offering possible next words. Their suggestions may be helpful because they know common expressions, but you remain the person deciding what to say. The analogy has limits: a predictive system ranks patterns in data and context; it does not necessarily share your purpose or know whether a suggested factual claim is correct.

Worked example

Suppose a toy test contains ten text prefixes, each with a writer's recorded intended next word. The system shows three suggestions for each prefix. The intended word appears first in five cases, second in two, third in one, and nowhere in the list in two.

Top-one accuracy is 5/10, or 50%. Top-three coverage is 8/10, or 80%. These numbers answer different questions: how often the first suggestion matches, and how often the intended word appears anywhere among the three options.

They do not show that users typed 80% faster or accepted 80% of suggestions. To assess benefit, a study must observe interaction time, corrections, and acceptance behavior. It should also recognize that the recorded next word is one intended continuation for this test, not the only grammatical continuation that could follow each prefix.

Illustrative case study

Fictional case: Maple Messages tests phrase suggestions for appointment replies. A suggestion ends with “I will attend,” even when the writer was preparing to decline. The wording is fluent, but accepting it would change the message's purpose.

The team keeps acceptance explicit and includes agreement, refusal, and uncertainty scenarios in evaluation. Reviewers examine the final sent-ready draft, not just whether the suggested phrase sounds natural. The case illustrates why useful assistance requires respect for intent. It does not claim that a particular keyboard makes this error or that a design change guarantees it will never occur.

Common misconceptions and mistakes

Do not assume the most likely next word is the most appropriate one. A common phrase may be unsuitable for the recipient, tone, or actual circumstances.

Do not equate high suggestion acceptance with high quality. People may accept text because it is convenient, and a longer suggestion may shape the content of their writing. Final-message review matters alongside interaction statistics.

Avoid assuming all predictive keyboards learn or store information in the same way. Personalization, processing location, and account settings are implementation-specific. A glossary definition cannot establish a product's data practices; those require checking the actual system and its documentation.

Business use

Predictive text can reduce repetitive entry and support people who find typing effortful. For workplace deployment, the relevant evidence is whether it helps users complete representative writing tasks accurately and comfortably.

A pilot can compare task time, correction effort, and perceived control with and without suggestions. It should include unfamiliar names and unusual terminology, where common-language predictions may be less helpful. Product teams should make it easy to ignore or reverse a suggestion. An assistant that saves keystrokes but creates accidental commitments or changes important details may fail the communication task despite strong model-level scores.

Check your understanding

A system includes the intended word among its three suggestions in eight of ten test cases. Does that establish 80% faster typing?

Answer: No. It establishes 80% top-three coverage under that test. A typing-speed claim requires observed timing and a suitable comparison.

Related terms

Language model
autocomplete
next-word prediction
autocorrection
text entry
personalization
assistive communication

Learn next

Compare top-one prediction accuracy, candidate-list coverage, and user acceptance. Then test a small set of writing tasks while recording both time and unintended changes to meaning, rather than treating fewer keystrokes as the only success measure.

Sources

Primary sources: Arnold, Gajos, and Kalai, On Suggesting Phrases vs. Predicting Words for Mobile Text Composition, introduction and interface discussion: https://www.microsoft.com/en-us/research/wp-content/uploads/2016/12/arnold16suggesting.pdf ; Hard and colleagues, Federated Learning for Mobile Keyboard Prediction, primary abstract: https://arxiv.org/abs/1811.03604 . Relevant primary descriptions were opened and read. The article makes no claim that all keyboards share these implementations. Test counts and the Maple Messages case are original illustrations.

Previous: Predictive state representation Next: Preference alignment