“Large language models can be co-opted by feeding them fake academic papers, as demonstrated by an experiment where a model diagnosed a user with a fabricated medical condition after the fake paper was posted online for three months.”