AI is boosting paper counts. The signal of quality is getting fuzzy.
Date: December 24, 2025 - Source: Cornell University
Large language models are helping researchers ship more manuscripts, faster. That lift is biggest for scientists writing in English as a second language. The tradeoff: more polished text that doesn't always match the scientific value underneath. Reviewers, editors, and funders are now sifting through more, with weaker cues.
What the Cornell team did
Researchers analyzed 2+ million preprints from January 2018 to June 2024 across arXiv, bioRxiv, and SSRN. They built a detector to estimate which authors were likely using LLMs for writing, then tracked posting rates before and after apparent adoption. They also checked which papers were later accepted by journals and studied how writing style related to acceptance.
What changed - by the numbers
- Output jump: roughly +33% more papers on arXiv for authors flagged as LLM users; 50%+ on bioRxiv and SSRN.
- Language advantage: researchers at Asian institutions posted 43.0% to 89.3% more after apparent LLM adoption, depending on the platform.
- Search behavior: AI search (notably Bing Chat) surfaced newer papers and relevant books more often than traditional tools, which skewed older and highly cited.
- Quality signal shift: for human-written papers, higher writing complexity correlated with journal acceptance. For AI-assisted papers, that link weakened or flipped-polished prose did not predict acceptance.
"It is a very widespread pattern, across different fields," said Yian Yin of Cornell Bowers CIS. First author Keigo Kusumegi noted that LLM-enabled search seems to connect researchers to more diverse sources, which could support more creative combinations of ideas.
Why this matters for researchers and writers
- Speed and access are up. Barriers from English fluency are lower.
- Review signals are noisier. Smooth writing is less informative about substance.
- Evaluation practices need updates. Publication counts and prose quality are weaker proxies for impact.
Practical guardrails for labs and authors
- Add an "AI use" statement: where LLMs were used (editing, summarizing, draft text, code comments), which model, and how outputs were verified.
- Separate polish from substance in your own checks: contribution, novelty, methods clarity, data/code availability, statistical power, and reproducibility.
- Keep an audit trail: original drafts, prompts, and major edits. It helps with integrity questions and internal reviews.
- Use LLMs for language editing and structure, not idea inflation. Don't let the tool over-claim or invent.
- Preprint wisely: share code/data, add a concise limitations section, and invite targeted feedback before journal submission.
- For non-native English speakers: draft in your native language, translate with an LLM, then back-translate a section to catch meaning drift.
For editors, reviewers, and funders
- De-emphasize writing complexity as a quality cue. Emphasize transparent methods, data/code, preregistration (when applicable), and effect sizes.
- Require AI-use disclosures and encourage structured abstracts with concrete claims, datasets, and evaluation metrics.
- Adopt triage checklists: contribution statement, comparison to baselines, error analysis, replication plan, and negative-result openness.
- Use automated tools to cluster and route submissions, but spot-check for AI artifacts and citation oddities.
- Reward openness: badges or priority for reproducible packages and registered reports.
Smarter literature search with AI (without getting misled)
- Use AI search to broaden the net, then verify citations in primary databases and publisher sites.
- Ask for "most recent and methodologically similar" work, and require the model to return full citations with DOIs. Check each DOI.
- Compare AI-found items to a traditional query. If overlap is low, merge both sets and de-duplicate.
- Maintain a short note in your methods on how you built the bibliography, including any AI tools used.
What's next
The findings are observational. The team plans controlled studies to test cause and effect, including randomized access to LLMs. A campus symposium is scheduled for March 3-5, 2026 to explore how generative AI is changing research and how policy should respond.
As Yin put it: the question isn't "did you use AI?" It's "how did you use it-and did it actually help?"
Study details
- Paper: "Scientific Production in the Era of Large Language Models" (published Dec. 18 in Science)
- Authors: Yian Yin; Keigo Kusumegi; Xinyu Yang; Paul Ginsparg; Mathijs de Vaan; Toby Stuart
- Support: National Science Foundation
Useful links
Want to skill up on responsible AI writing?
If you're formalizing lab or editorial practices-or you write for a living-practical training helps. See focused resources on prompts, editing workflows, and audit trails here: Prompt engineering guides and courses.
Your membership also unlocks:
