A paper "A Ship of Theseus: Curious Cases of Paraphrasing in LLM-Generated Texts" produced by the researchers from KInIT, the MIT Lincoln Laboratory, the University of Pennsylvania and Indiana University in the US investigates how repeated paraphrasing by Large Language Models (LLMs) affects the authorship of a text.
With the rising mainstream use of LLMs such as ChatGPT, research in AI text generation and detection has expanded significantly, where paraphrasing plays a crucial role.
Drawing inspiration from the Ship of Theseus paradox, the study frames each paraphrasing iteration as a replacement of linguistic “planks,” showing that while the core content remains intact, the writing style shifts progressively away from the original author’s unique expression. This stylistic divergence leads to declining performance in authorship attribution models, raising critical questions about originality and authorship in AI-assisted writing.
Major findings include:
1. Style deviates a lot more than content after paraphrasing:
Paraphrasing with LLMs preserves the original semantics but leads to substantial style deviation, which is why classifiers fail to correctly attribute the paraphrased versions.
2. LLM paraphrasers deviate the style to the LLM style model:
LLM paraphrasers align the text’s style with the LLM’s style model, often making paraphrased text more similar in style to the LLM than to the original author. This results in misclassification of the paraphrased text as LLM-generated.
3. Subsequent paraphrasing differs for LLM and PM paraphrasers:
There is a notable performance drop after the first paraphrasing, but only a 3–4% average decrease from the second paraphrase onward. PM paraphrasers show a consistent decline in style compared to LLM paraphrasers across subsequent versions.
4. The choice of paraphraser impacts performance and style deviation:
ChatGPT shows stronger paraphrasing than PaLM2 in both performance drop and style deviation. Lexical diversity also plays a role, as shown by Dipper’s varying impact under low, moderate, and high settings.
5.Performance varies across datasets and sources:
Paraphrasing has a milder effect on formal datasets like XSum and SciGen due to consistent style. Informal datasets (e.g., ELI5, CMV, Yelp) exhibit greater style variance and are more affected. Texts by Humans and ChatGPT retain style better through paraphrasing than those from PaLM2 and Tsinghua.
These findings suggest that authorship should be task-dependent. Given the increasing prevalence of LLMs in generating and enhancing text, the research can provide a sound basis for addressing plagiarism and copyright disputes in the future involving LLMs.
Link to Zenodo: https://zenodo.org/records/13851398