The Erosion of LLM Signatures: Can We Still Distinguish Human and LLM-Generated Scientific Ideas After Iterative Paraphrasing?

2025-12-04

AI Summary

This study systematically investigates, for the first time, the problem of distinguishing human-authored from LLM-generated scientific ideas—a critical challenge in scientific creativity attribution. We evaluate state-of-the-art detection models on scientific ideas subjected to multiple rounds of human paraphrasing. Results show that five successive paraphrasing iterations substantially degrade detection performance, reducing average accuracy by 25.4%; notably, non-expert, simplified paraphrasing styles prove most confounding. Incorporating the original research question as contextual prompt improves detection accuracy by up to 2.97%. The findings reveal that semantic-preserving yet representation-shifting paraphrasing poses a fundamental challenge to LLM provenance tracing, exposing the limited robustness of current detectors in scientifically grounded, cognitively nuanced scenarios. This work provides novel empirical evidence on the cognitive boundaries of LLM-assisted scientific ideation and underscores the need for more semantically resilient attribution methods.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Cognitive Modeling & Cognitive Systems: Computational Creativity

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

With the increasing reliance on LLMs as research agents, distinguishing between LLM and human-generated ideas has become crucial for understanding the cognitive nuances of LLMs' research capabilities. While detecting LLM-generated text has been extensively studied, distinguishing human vs LLM-generated scientific idea remains an unexplored area. In this work, we systematically evaluate the ability of state-of-the-art (SOTA) machine learning models to differentiate between human and LLM-generated ideas, particularly after successive paraphrasing stages. Our findings highlight the challenges SOTA models face in source attribution, with detection performance declining by an average of 25.4% after five consecutive paraphrasing stages. Additionally, we demonstrate that incorporating the research problem as contextual information improves detection performance by up to 2.97%. Notably, our analysis reveals that detection algorithms struggle significantly when ideas are paraphrased into a simplified, non-expert style, contributing the most to the erosion of distinguishable LLM signatures.

Problem

• Detect LLM vs human-generated scientific ideas after paraphrasing
• Evaluate SOTA models' ability to distinguish idea sources
• Assess impact of iterative paraphrasing on detection performance

Innovation

• Evaluating SOTA models for idea source attribution
• Incorporating research context to improve detection performance
• Analyzing paraphrasing impact on LLM signature erosion