Plagiarism in AI Generated Research

Large language models (LLMs) are now being promoted as tools that can sketch out bold research directions, generate full proposals, and even draft papers that look publication ready. These claims naturally raise the question about how original such outputs are. In a recent study, Gupta and Pruthi examined this closely and found that a significant number of artificial intelligence (AI)-generated ideas borrow heavily from existing work.

In this work, thirteen natural language processing (NLP) researchers were asked to review 50 AI-generated proposals and papers. Unlike earlier evaluations that simply rate novelty or feasibility, the reviewers were asked to approach each document with suspicion and look for potential overlap with prior research. Their assessments revealed that roughly one quarter of the documents reused key parts of earlier methods, often with only cosmetic changes in phrasing.

What makes this especially worrying is that the same documents had already been screened by several plagiarism-detection pipelines, including Semantic Scholar-based search tools, LLM-driven similarity checks, and commercial systems. None of these filters raised alarms. Some of the flagged documents had even been highlighted elsewhere as examples of ‘novel’ LLM-generated research. Together, these observations show that existing checks are easy to circumvent and that any AI-generated research content must be examined with care before being trusted. Given the gravity of these findings, this work was conferred the Outstanding Paper Award at ACL 2025.