ManBert

The rapid adoption of large language models (LLMs) such as ChatGPT and Gemini has raised important questions about their impact on academic writing practices. In this study, we examine whether AI-assisted manuscript writing has increased following the widespread availability of generative AI tools. We analyze papers from 2021, 2023, and 2025, focusing on authors who published in all three years to enable a controlled comparison over time. We develop an AI detection model trained on a curated dataset of 1,024 samples, split into training, validation, and test sets. The model achieves an F1 score of 0.95, outperforming existing approaches such as ZeroGPT, GPTZero, Sapling.ai, and BERT-based baselines. Applied to full-text articles (~1,000,000), our method provides a more comprehensive assessment of AI-assisted content than prior abstract-based analyses. Our results show a statistically significant increase in AI-assisted writing from 2021 to 2025. This trend holds consistently across regions (North America, South America, Europe, and Asia), across disciplines (e.g., life sciences and engineering), and within both STEM and non-STEM domains in each region. These findings demonstrate that AI-assisted writing is rapidly expanding across time, geography, and fields, highlighting its growing role in scholarly communication.

Results

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6178360