- AI tools increased researchers' paper output by 36 to 60 percent.
- Non-native English speakers saw the largest gains, up to 89 percent more papers.
- AI-generated language complexity now inversely predicts publication success.
Scientists using AI are cranking out dramatically more research papers, but the quality screening methods that academic publishing relies on are breaking down.
According to a new study in Science, AI has triggered a surge in productivity while simultaneously making it harder to separate genuine scholarship from what researchers call "scientific AI slop."
In an article in The Conversation, Vitomir Kovanovic explains how researchers from UC Berkeley and Cornell University analyzed over a million preprint articles from 2018 to 2024 and found something unexpected: AI tools are changing not just how much scientists write, but also challenges whether traditional markers of quality still mean anything.
Key figure
36–60%
increase in paper output when scientists use AI writing tools
What is AI slop?
"AI slop" refers to low-quality, generic content mass-produced by AI with little or no human oversight, editing, or creative input. The term draws an analogy to cheap, unappetizing food – content that technically exists but lacks originality, accuracy, or any authentic human voice. At Science Reader we try to be a counter-weight by using AI responsibly in our content workflows.
The Productivity Explosion
When researchers started using AI, their output jumped between 36% and 60% depending on the platform.
The increase was most dramatic for non-native English speakers, particularly Asian authors, whose productivity soared by 43% to 89%. Authors from English-speaking institutions saw more modest gains of 24% to 46%.
The pattern suggests AI serves as a language equalizer, helping scientists whose first language isn't English compete on more equal footing. That's genuinely promising.
When Complex Language Hides Weak Science
Here's where things get troubling. The Berkeley and Cornell teams discovered that AI-generated articles used more complex language on average. In traditional academic writing, complex language correlated with higher publication rates, a signal of quality and rigor.
But for AI-assisted papers, that relationship flipped. The more complex the AI-generated language, the less likely the paper was to be published. What makes this significant is that it reveals how AI linguistic sophistication can mask poor scholarship rather than enhance it.
Journal editors and reviewers who rely on writing quality as a quick screening tool are increasingly being fooled. The usual heuristics for spotting strong research are failing.
Editor's note: At Science Reader we use AI tools in our editorial workflow, but with human verification at every stage, quality mechanisms and editorial oversight. We check sources, verify attributions, cowrite the content, and review for epistemic accuracy - the opposite of the "minimal human effort" approach this article warns about.
AI Search Shows One Bright Spot
The study included an intriguing side finding about how AI-enhanced search affects what scientists read. Comparing Microsoft's AI-powered Bing Chat to Google's traditional search, researchers found Bing users encountered more diverse sources and more recent publications.
This challenges fears that AI search would trap scientists in echo chambers of widely-cited older papers. The technique behind this, retrieval-augmented generation, combines real-time search results with AI prompting.
Related reading
The One Thing AI Image Generators Can't Do
They generate endlessly but can't judge what's interesting. New research shows why your taste still matters.
→How To Fight The AI Slop Battle
AI is now embedded in word processors, email apps, image generators and spreadsheets. Avoiding it will soon be impossible whether researchers want to or not.
The immediate challenge is developing better evaluation methods. Quick assessments based on language sophistication no longer work. Kovanovic suggests journals will need deeper peer review focused on methodology and actual contributions rather than prose quality.
Can academic publishing adapt fast enough?
Some researchers are proposing fighting fire with fire, by using AI review tools like the one recently published by Stanford's Andrew Ng. Given the flood of submissions already overwhelming journal editors, automated screening may be the only viable option.
The larger question remains unanswered: can academic publishing adapt fast enough to separate genuine discoveries from the rising tide of sophisticated-sounding slop?
Sources / References
- Primary source: What the hyperproduction of AI slop is doing to science (The Conversation)
- Context Sources
