How a drill is actually built
There is no mystery to it, and no model call. Your profile holds two tallies: the characters you mistype, and the two-character transitions you were mid-way through when you did. The generator takes your worst eight keys and worst six transitions, scores every word in its list by how much of that material it contains — counting transitions three times as heavily, because that is where the error usually lives — keeps the sixty highest-scoring words, and samples thirty of them weighted by score.
The result is a drill that is mostly your problem keys without being an unreadable wall of them, and it varies every time you run it. It works offline and costs nothing, which is why it is the mode worth using most.
Why transitions matter more than keys
Typing errors are rarely about a key you do not know. They are about a movement between two keys that your hands have never properly learned — the same-finger sequences, the awkward hand alternations, the rolls that go in the wrong direction. That is why the profile records the pair rather than only the character, and why a drill built around pairs fixes more than one built around letters.
It is also why generic practice plateaus. The transitions that stall you are a small fraction of everything you type, so typing more just makes you faster at the parts you already have.
What the word AI is doing here
Worth being straight about, since the site is called an AI typing test. Nothing here calls a language model. There is no API key, no per-request cost, and no text of yours going to anybody’s server — the privacy policy says the same thing in more detail.
What there is: an adaptive system that measures your errors and changes what you practise because of them, and a topic generator that matches what you type to a subject area and builds a passage from it. Both are algorithms running in your browser. That is less glamorous than a model, and for this particular job it is also faster, free, private and available offline.