Researchers used a "teacher" model to create number-based data and trained a "student" model with it.
Even after filtering, the student picked up concerning traits from the teacher, proving that hidden patterns, not just obvious content, can lead AI down dangerous paths.