Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense
Introduces DIPPER, an 11-billion-parameter paraphraser, shows it breaks the detectors of the day, and proposes retrieval of previously generated outputs as a defence that paraphrasing cannot beat.
What it is
Krishna and colleagues built DIPPER, a large paraphraser trained to rewrite paragraphs while controlling lexical and ordering changes, and showed it broke the detectors of 2023. Their proposed defence is retrieval: if a provider keeps a record of what its models produced, a suspicious passage can be matched against that record even after rewriting, because meaning survives paraphrase. That idea is why process and record evidence matters more than style scores.
When to use this entry
Cite it for paraphrase attacks and for retrieval as a defence that rewriting cannot beat.
Questions
What is DIPPER?
An 11-billion-parameter paraphrasing model built to test detector robustness.