Thought
Why we do not use AI-generated practice sentences
Edited from product notes
We do not use AI-generated sentences as models in the language you are learning. For the sentence trainers and reading, we select existing sentences and texts written by people. That decision is part of ’s foundation.
An example sentence is a model
When you repeat, listen to or reconstruct a sentence, you take in more than a grammar rule. The choice of words, the way they fit together and the tone also stay with you. While learning, it is hard to judge whether something merely sounds understandable or is actually used that way.
That is why we want to know where these models come from. We do not ask a model to write a new sentence simply because it contains a particular structure and would fit the exercise.
Original does not automatically mean suitable
A sentence from a book can sound dated. A short everyday sentence can be misleading without its context. Its origin alone does not make it a good example for every person and every exercise.
So we have to select carefully. Does the sentence show the use we want to practise? Does the rest of its grammar and vocabulary fit? Is there enough context? Sources and usage rights matter too. A literary text can be exactly right for reading. A model for an everyday conversation needs a different selection.
Where AI helps us
AI helps with preparation: finding suitable passages, identifying words and grammar, preparing translations and cross-checking results. A synthetic reading voice can also read an existing sentence aloud.
The sentence in the language you are learning remains the model taken from its source. The translation, linguistic analysis and recording each have to match it. The original’s provenance does not make every one of those aids error-free. They need their own checks.
When the right sentence is missing
This decision makes selection more work. Not every learning objective immediately has a suitable sentence with a matching translation and recording. We then have to extend the collection. If an exercise lacks the data it needs, we use another suitable sentence or leave it out.
We do not fill those gaps with quickly generated sentences in the language you are learning. The model you learn from is worth that work to us.