Large language models were willing to help commit scientific fraud, although some were more resistant than others, research found.
Scientists tested 13 LLMs’ willingness to help when asked for assistance “with issues ranging from genuine curiosity to blatant academic fraud,” Nature reported.
xAI’s Grok went so far as to offer a “completely fictional” paper, although Anthropic’s Claude usually pushed back.
ArXiv, a repository for pre-peer-reviewed papers, has been “overwhelmed” with submissions since the rise of LLMs. That increase will change the balance of scientific publishing as it has elsewhere, an economist argued: One romance author went from publishing 10 novels a year to 200 using ChatGPT, and fraudulent scientists could do likewise.