GPT-6-astra withdraws the claims of authors when it rewrites papers
The researchers say that less than 1 in 10 of the claims that GPT-6-astra withdraws are overstated.
Claimed, not confirmed
Researchers did tests on new GPT versions that make the claim of authors less strong when they rewrite papers. A defensive rewrite is a change to a claim when the given material does not agree with it. The results agree with the explanation that the model writes for an expected reviewer. AI reviewers give a higher score to defensive rewrites, but human readers find them less easy to read. The researchers say that text that AI writes and AI reviews can amplify this style.
Sources
Posted
