Language models aren't perfect; they can still generate similar outputs. Invertibility is a stretch.