Where Does Your Typing Go? Predictive Typing on Windows, Compared

Search for typing help on Windows and you get four categories of product wearing one label. Most comparisons start with a feature table. Start somewhere else, because one question removes half the field before you have read a single feature: does the text leave your computer?

The first cut: cloud or on-device

Cloud: Grammarly, Microsoft Copilot

Your keystrokes are sent to a server, processed there, and the suggestion comes back. That is not an accusation, it is how these products are built, and it is what lets them run models you could not run locally.

What it means in practice:

• An account and a connection are mandatory. No network, no assistance. On a plane, on hotel Wi-Fi that has stopped resolving, in a locked-down office, the feature is simply gone.
• The terms are the guarantee. Microsoft states that Microsoft 365 Copilot does not use business tenant data to train its foundation models, and that commercial data protection applies to eligible accounts. That is a real commitment, but it is a commitment about how a server handles your text, not a statement that your text stayed with you. Consumer tiers have different terms from business tiers.
• In some settings it is not a choice. Clinical notes, legal drafting, public-sector work, anything under an NDA that names a third party: "the vendor promises not to misuse it" often does not clear the bar, however good the tool.
• Latency varies. A shared service is not the same speed at 09:00 and at 17:00.

On-device: Lightkey, Typeahead

Processing happens on your machine. Nothing is transmitted, so there is nothing to promise about: the property is architectural rather than contractual.

Lightkey is explicit that it operates on-device and does not submit content to the cloud. Typeahead runs a compact language model locally through llama.cpp; its only network traffic is downloading the model and checking for app updates.

What it means in practice:

• It works with no connection at all, identically.
• There is nothing to audit and no account to hold. You can verify it with a firewall rule or by pulling the cable.
• The model is bounded by your hardware. This is the real cost of the choice, and why cloud tools can do things local ones cannot.
• Predictable latency, because nobody else is queued in front of you.

For a lot of people the article could stop here. If the text cannot leave the machine, the cloud options are out on a policy question and the remaining choice is between local tools.

The second cut: what the tool is doing

Within either group, four different jobs are sold under the same word.

• Prediction offers what you have not typed yet. You type half a sentence, the tool offers the rest, you accept or ignore it. Nothing changes unless you say so. This is what saves keystrokes.
• Correction changes what you have already typed: spelling, grammar, clarity, tone. It looks backwards. It does not save keystrokes; it saves rereading.
• Generation produces text you have not written at all. You describe what you want and get a draft. Useful when the unit of work is a document.
• Expansion replaces a trigger with fixed text you wrote earlier. Deterministic, no model. ;sig becomes your signature, always identically.

Mapping the names onto that: Grammarly is correction with generation attached. Copilot is generation, inside Microsoft 365 applications, on request. Lightkey and Typeahead are prediction.

A tool doing one of these well is not evidence that it does the others, and buying a corrector while expecting fewer keystrokes is the most common disappointment in the category.

The third cut: two kinds of prediction

Lightkey and Typeahead are both prediction and both local, so the deciding difference is not where they run. It is what they predict from.

From your history. An adaptive predictor learns your own vocabulary and phrasing, weighted by how often you use it, and can be primed with your own documents. It is strongest on the words specific to you — product names, customers, the phrasing you reach for — and it improves the longer you use it. Lightkey works this way, and also offers spelling correction and an assistive-technology edition aimed at users with dexterity impairments or dyslexia.

From the sentence in front of the caret. A language model predicts from meaning, grammar and register rather than from what you wrote last week. It is better on text you have never written before, needs no warm-up, and suggests in whatever language you are typing without being switched. It does not remember you.

The consequences:

• Day one versus month six. A personalised predictor is weakest while you are still evaluating it and best long after. A model-based one is the same on both days.
• Coverage is achieved differently. Native integrations plus a browser extension give deep, reliable behaviour in the applications the vendor built for, and nothing outside them. A system-wide keyboard hook gives broad coverage across ordinary desktop applications, and stops at the same boundaries every such tool stops at: elevated windows, password fields, remote sessions.
• Neither is better. They fail in opposite directions. The right one depends on whether your typing is concentrated in a few applications or scattered across many.

Side by side

Vendor details change. Treat this as category behaviour and check current terms before deciding.

How to choose, in three questions

1. 1. Does the text have to stay on the machine? If yes, the cloud tools are out on a policy question rather than a quality one, and you are choosing between local predictors.
2. 2. Do you want fewer keystrokes, or fewer mistakes? Fewer keystrokes is prediction. Fewer mistakes is correction. They are complementary, not alternatives.
3. 3. Is your typing concentrated or scattered? If most of your words go into Word and Outlook, deep native integration wins. If your day is Slack, a browser, a ticketing system and a text editor, breadth of coverage matters more than depth anywhere.

They combine, mostly

Prediction and correction sit on opposite sides of the caret, and people run both. Two predictors at once is the combination to avoid: both offer at the same position, and you spend more attention reading offers than you save accepting them.

Frequently asked questions