"Over a year, we collected 4,900 summaries. When we analysed them, we found that six of ten models systematically exaggerated claims they found in the original texts"

So it turns out llms trained largely on Internet science articles make the same mistakes as are made by science journalists.