Alpha-OSK's prediction engine avoids about half of all keystrokes when it is tested offline, replaying old sentences back through the model. That number assumes a typist who takes every suggestion the instant it appears, and never a moment later. Nobody knows yet how much of that saving survives contact with an actual person, deciding in real time whether a pill is worth the click. That gap between the lab number and a lived one is what this study measures.
That is not unusual for the field, and it is exactly why an outside check is worth running.
Most claims about assistive-keyboard prediction, including the ones on this site, come from the team that wrote the software, measured against sentences nobody actually typed. That is the normal way this kind of software gets evaluated, and it is also why the number that gets repeated is rarely the number anyone else has checked.
Alpha-OSK has exactly one person who has used it for real work over real months: Owen, who also wrote the prediction engine it runs on. A study with outside participants is the first chance to see whether what works for him works for anyone else, typing under different conditions with a different pointer and a different vocabulary.
For this keyboard, and for the other assistive tools that use the same approach, because the protocol and the anonymised data are published for anyone to check, not just the headline number.
If the saving turns out to be smaller with a real person than offline testing predicts, that is worth knowing too. It says where the engine's effort should go next, rather than where a lab test already said it succeeded.
Both are optional. Both can be turned off or stopped at any time. Neither needs an account.
Never sent: the words you type, any other content, your IP address, or key frequencies.
Off by default. Turn it on in Settings → Data & Privacy → Privacy, or accept the invitation the installer shows on first run. You can delete everything you have contributed, at any time, with one button, and opting out discards the ID itself, so a later opt-in can never be linked back to what came before it.
Read the privacy sectionYou copy short phrases while the keyboard measures typing speed and accuracy, once with prediction on and once with it off, in a randomised order. You can stop at any point, and if you ask, your partial data is discarded rather than kept.
Not open for enrolment yetEvery session produces the same four measures, with prediction on and with it off, so each participant is compared against themselves.
Words per minute: how fast the target text lands on screen, with prediction on and with it off.
The share of characters that end up wrong, counting what a backspace had to fix along the way.
How many clicks it actually took to produce each character of the target phrase: the real cost of a keyboard you operate by pointing rather than pressing.
How many of those clicks the predictions actually removed, measured against a real person rather than a replayed sentence. Offline testing puts this at about half. Whether that holds up here is the reason the study exists.
The same rule as the rest of the site: prediction runs on your machine, and the study asks for no more than either option above already describes.
Unless you switch sharing on, nothing about your typing or your study session leaves your computer.
There is nothing to register, for the keyboard or for either kind of participation.
The anonymous ID is a random number with no link to your name, your email, or your device.
No name, email, or address is attached to anything you type during a session.
At any point, for either kind of participation, without needing to explain why.
The full protocol and consent form are published before anyone enrols, not written up afterwards to match what happened.
This section is for anyone deciding whether to trust the numbers this produces: reviewers, clinicians, and other developers of assistive input tools.
This study is run by the project's author. It has no institutional affiliation and no ethics board (IRB) review. The protocol and consent form below are published in advance, in place of that oversight, and the data and the analysis will be made public once the study runs, so the result can be checked rather than taken on trust. Being straight about this is the point of this section.
Read the source documents:
Alpha-OSK