How we check the grammar we teach you

KorrektWort has a grammar trainer. One German sentence, one gap, three options, and exactly one of them is right. After you answer, one line tells you why.

Those exercises are written with AI. You should know that, and you should also know what we do about the obvious problem with it, because you are going to trust what that one line tells you about German.

The problem, in one example

Here is a real exercise from an early version. The sentence was Ich habe mein Buch _____ Hause vergessen, and the correct answer was zu. That is right. A learner would have chosen zu, seen a green tick, and felt good.

Then the explanation said that Hause is a dative article. There is no such thing. Hause is an old form of the noun. The exercise worked perfectly and quietly taught something false about German.

That is the danger with AI-written learning material, and it is not the danger most people expect. The answers are usually right. It is the explanations that go wrong, and they go wrong in fluent, confident, textbook-sounding sentences. If you read only the exercise, you would never notice.

So we measured it instead of trusting it

We generated a batch of exercises and had them graded strictly, one by one, without knowing which had been changed or tuned. In that early version, 15% contained something factually wrong, and in another 13% a second option was quietly also correct, which would have marked you wrong for writing correct German.

Not one of those exercises was ever shown to a learner. They existed only in testing. That is the entire reason for testing.

What an exercise has to survive now

Before any gap reaches you, it has to pass every one of these, and any single failure means the exercise is deleted rather than repaired.

• The sentence itself must be natural, correct German.
• The explanation is written separately, by a different step that does not know how the exercise was built, so it has to describe the actual rule instead of defending a choice already made.
• Every claim in the explanation is fact checked, and every German word it quotes must really appear in the sentence.
• One step exists only to attack the exercise: it takes each wrong option, puts it in the gap, and argues that it also works. If it succeeds even once, the exercise is gone.
• Finally two independent panels, each reading the exercise from three different angles, have to agree completely. If one reader out of six hesitates, it is deleted.

Roughly half of everything we generate dies somewhere in there. That is fine. The sentences are cheap and your trust is not.

The part we cannot promise

Everything above is still machines checking machines. They can share the same blind spot, and if they all misunderstand the same piece of German, it slips through and nobody notices.

There is exactly one thing that catches that, and it is you. Every exercise has a small This looks wrong button. Press it whenever something does not match what you know, even if you are not sure, even if you cannot explain why. Nothing bad happens, you do not lose anything, and a human reads every report. If there is any doubt at all, that exercise leaves the app immediately.

Our promise is simple. If you write correct German, KorrektWort must never tell you that you are wrong. When it does, we would rather hear about it than not know.

Why we are telling you this

Plenty of apps generate their content with AI now and none of them mention it. Saying nothing would have been easier for us too.

But you are learning German for a reason, maybe an exam, maybe a job, maybe a life you are building here in a language that is not yours. You are going to remember what we tell you. That deserves to be checked, and it deserves to be checked out loud.