More of the same? Are scientific papers losing originality?

Ilan Doron-Arad, Elchanan MosselarXiv 2026 · arXiv

TL;DR

Is the explosion of scientific publishing actually producing a broader frontier of ideas, or just "more of the same"? A new study from MIT researchers finds that scientific papers have become significantly less textually distinct since 2021, with the share of text resembling recent work more than tripling by 2025. This trend isn't driven by a few "copy-cats" but by a widespread homogenization of language and ideas across the entire scientific literature.

What Problem Does This Paper Address?

Scientific output is growing at a breakneck pace, doubling roughly every 17 years. While this suggests a thriving community, it raises a fundamental question: is originality keeping pace with volume?

Measuring "originality" is notoriously difficult because it’s multidimensional. Previous attempts have looked at citation networks or the introduction of new keywords, but these often miss the nuance of how ideas are actually framed and written. The authors of this paper tackle a specific, measurable facet of originality: textual distinctiveness. They wanted to know if full-text papers are becoming increasingly semantically similar to the work that immediately preceded them, and if so, what is driving that change.

Methods

The researchers developed a sophisticated semantic similarity pipeline to screen over 26,000 full-text papers.

  1. Semantic Embedding: They used a modern language model (all-mpnet-base-v2) to convert 160-token passages into 768-dimensional numerical vectors. This allows the system to detect similarity in meaning, even if the specific words differ.
  2. The "Fixed-Size" Screen: To avoid the bias of a growing literature (where a newer paper has more "targets" to accidentally resemble), they compared every paper against a fixed-size sample of 3,000 papers from the preceding five years.
  3. Matching Criteria: A match was defined using conservative thresholds: high cosine similarity ($\ge 0.84$), at least six shared terms, and—crucially—no exact sequences of five words to filter out simple quotes or plagiarism.
  4. Idea-Level Analysis: They used a Large Language Model (LLM) to summarize matched passages and classify whether they expressed the "same scientific proposition," moving beyond mere phrasing to the underlying ideas.
The semantic similarity pipeline: from passage embedding to vector comparison.

Datasets

  • CORE Database: The primary source, covering eight diverse fields (astrophysics, bioinformatics, chemistry, climate change, machine learning, medicine, psychology, and quantum computing) from 2015 to 2025.
  • Europe PMC: An independent database used to successfully replicate the findings.
  • arXiv: Used for a qualitative robustness check in computer science sub-fields.

Key Insights and Ideas

  • The 2021 Inflection Point: Textual resemblance was remarkably stable from 2015 to 2020. Starting in 2021, the trend broke sharply upward. This timing coincides with the rapid adoption of LLM-assisted writing tools in academia.
  • Homogenization, Not Plagiarism: The increase wasn't caused by a few papers borrowing heavily from others. Instead, the "resemblance" spread across the literature. More papers now contain at least one passage that echoes recent work.
  • The Citation Gap: Surprisingly, nearly all of this growth occurred among pairs of papers that do not cite each other. Only about 6.4% of matched pairs had a detected citation link. This suggests a convergence toward a "shared disciplinary language" rather than intentional building on specific previous results.
  • Idea Convergence: The trend isn't just about "boilerplate" methods text. The number of pairs expressing the same scientific idea increased more than five-fold in 2025 compared to the pre-2021 baseline.

Key Results

The quantitative shift is stark. The "similarity coverage" (the fraction of a paper's text matched to prior work) rose from a stable 0.432% (2015–2020) to 1.343% in 2025—a 3.11-fold increase.

Trends in semantic similarity: (a) mean text matched, (b) share of papers with matches, and (d) a heatmap of similarity by publication lag.

As seen in Figure 1b, the share of papers containing at least one matched passage jumped from roughly 9% to over 25%. The researchers also found that this pattern held across all eight scientific fields studied, though the magnitude varied.

Limitations and Future Work

The authors are careful to note that semantic similarity does not equal misconduct. A decline in textual distinctiveness might signal a broader decline in originality, but it could also reflect a convergence in how scientists are taught to frame their work.

  • Causality: While the timing points toward LLMs, the study does not definitively prove they are the cause. Other factors, like electronic access narrowing attention to a smaller set of "canonical" papers, could play a role.
  • Conceptual Novelty: The study measures similarity at the passage level. It does not assess the novelty of a paper’s overall contribution or its experimental results.
  • Future Directions: The authors suggest that tools making "originality visible" could help reviewers and funders identify truly distinct work. However, they warn against turning "distinctiveness scores" into a target, as this would likely lead to researchers gaming the system with intentionally bizarre phrasing.

Summary updated

Abstract

Scientific output is growing rapidly, but it is unclear whether the expanding literature remains original or is increasingly repeating itself. Originality has many dimensions. One dimension can now be measured directly: how textually distinct a paper is from the work that came before it. Using semantic language models, we screened more than 26,000 full-text papers across eight fields in the CORE database (2015--2025) for passages resembling recent literature. We compared every paper against a fixed-size sample of earlier work, so increases in resemblance are not a by-product of literature growth. We find that the share of a paper's text resembling recent work was stable before 2021, then rose sharply, more than tripling by 2025. The increase came not from a few papers borrowing more heavily but from resemblance spreading across the literature: more papers now contain passages that echo recent work, and almost none of this growth is explained by cited work. We also found an analogous increase in passages expressing the same scientific idea. The pattern held across robustness checks and replicated in a second, independent database. Along this one measurable axis of originality, scientific writing has become less distinct from the work preceding it.
BibTeX
@misc{doronarad2026more,
  author = {Ilan Doron-Arad and Elchanan Mossel},
  title = {{More of the same? Are scientific papers losing originality?}},
  year = {2026},
  eprint = {2609.32963},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.32963},
}
Figure