The hidden liability in AI education is not that institutions are collecting data. It is that their systems may already be manufacturing judgments they cannot easily justify