4 ms·

> For all the complaints about AI generated content showing up in scientific journals, I'm exited for the flip side, where an LLM can review massive quantities

• themanmaran 2y ago
  Not in a zero shot approach. But LLMs are more than capable of solving a similar scenario to the one presented: - Parse all papers you want to audit - Extract images (non AI) - Diff images (non AI) - Pull captions / related text near each image (LLM) - For each image > 99% similarity, use LLM to classify if conclusions are different (i.e. highly_similar, similar, highly_dissimilar). Then aggregate the results. It wouldn't prove fraud, but could definitely highlight areas for review. i.e. "This chart was used in 5 different papers with dissimilar conclusions"