What counts as “AI generated” here? It’s likely many Nature papers currently secretly have sentences which were originally output or revised by LLMs.

@JasonQ Only papers generated in whole by an LLM, not in part.

@ChaseStevens How will this be verified? I'm sure there are probably some researchers using LLMs to revise the wording in their paper by rewriting it totally without disclosing that

@spiderduckpig almost no papers are text from a single person and there’s always changes in peer review so actually I’d be very sure of the opposite, there’s almost no way for a paper to go through being totally rewritten by AI. Someone is going to want something changed from the wording/content even if it’s not objectively any better. Co-authors, reviewers and editors are guaranteed to have personal preferences.

And that’s in the best case scenario, LLMs don’t do well getting technical details right at the edge of science IME because any novel publication deviates from the training/searchable corpus by definition. So an AI rewritten paper has very high likelihood of introducing mistakes by approximating text from similar but different pre-existing work.

The date on this is wrong

How much prompting is allowed for this to still resolve YES? If human researchers give the AI their notes to write into a paper, does that count?

@HarlanStewart Yes, your example would count.

Does this include only the flagship journal, or also things like "Nature Physics"?

@ScottLawrence Only the flagship journal.