Most alarmingly, after four months of reliance on ChatGPT, participants wrote worse once it was removed than those who had never used it at all.

Eighty-three percent of heavy AI users couldn’t recall key points from what they’d “written,” compared to only 10 percent of those who composed unaided. Neutral reviewers described the AI-assisted writing as “soulless, empty, lacking individuality.”

the effects of BCG IV were initially done in the 1970s. practically speaking that would be a pretty tough route to sell. the excitement is over having a model of protection, understanding how it works and developing a vaccine out of that knowledge.

thanks Oded; that is reassuring! I think that could be clarified in the FAQ section "Deleting a manuscript immediately ends qed’s licence to use it..Any model parameters already learned or anonymized stats remain because they cannot technically be rolled back, but they contain no identifiable text."

this sounds cool. question: if someone uploads and later deletes a manuscript but it has already been used for training how do you know what the model has been trained on? is there any validation of the submission (what happens if someone writes a terrible paper or provides false data?)