AI is flooding science with papers, but many lack real substance

A new study from UC Berkeley and Cornell University reveals a paradox at the heart of AI's arrival in academic research: the technology is making scientists dramatically more productive, while simultaneously degrading the quality of what gets published.

Mathijs De Vaan, an associate professor at UC Berkeley Haas, watched this unfold firsthand. When ChatGPT and similar language models emerged in late 2022, he saw an obvious benefit—they could help researchers polish manuscripts and overcome language barriers that had long disadvantaged non-native English speakers. Writing clarity finally didn't have to be a proxy for scientific rigor.

But as these tools moved from novelty to standard practice across academia, a troubling question emerged: Was AI simply making research easier, or was it fundamentally breaking how science works?

The Productivity Boom

The numbers are striking. Scientists who adopted large language models saw their manuscript output jump more than 50% on bioRxiv and SSRN, and over one-third on arXiv. The effect was most pronounced among researchers whose first language isn't English—those with Asian names at Asian institutions experienced productivity gains approaching 90% in biology and social sciences. By contrast, Western researchers at English-speaking institutions saw more modest but still significant increases of 24% to 46%.

On the surface, this looks like a genuine win. For decades, the language barrier had been a real disadvantage for talented scientists working outside the English-speaking world. AI tools offered a way to level that playing field, letting the quality of ideas matter more than the polish of prose.

De Vaan himself noted another advantage: AI-powered search tools like Bing Chat are genuinely better at surfacing newer, more relevant research than traditional search engines, which tend to keep recycling the same canonical papers. "LLMs allow us to search a broader base and delve much deeper," he says.

Where It Falls Apart

Then comes the catch. For generations, writing quality has been an imperfect but reliable signal of scientific rigor. A paper that articulates complex ideas with precision suggests the researchers understand their subject deeply. These well-written papers were more likely to pass peer review and get cited frequently.

AI has inverted that relationship entirely. The study found that manuscripts polished by AI assistants show a striking reversal: the more complex and sophisticated the writing, the less likely the paper was to be published in peer-reviewed venues. The sophisticated prose was masking weak science, not signaling strong research.

"The robots now write more complex and sophisticated science than many human scientists," says Toby Stuart, another UC Berkeley Haas professor involved in the research. "But what our analysis shows is that scientific articles that were mostly automated are of substantially lower quality than human-written papers."

This creates an immediate crisis for the scientific community. Peer reviewers are already overwhelmed. The flood of plausible-looking but substantively weak papers threatens to clog the entire system, making it harder for genuine breakthroughs to get attention.

What Comes Next

The researchers suggest that AI itself might offer part of the solution—specialized "reviewer agents" that filter papers before they reach human reviewers. But they're clear that technology alone won't fix this. The scientific community needs to rethink how it evaluates research, funds work, and verifies claims.

De Vaan puts it plainly: "The institutions that begin experimenting now with new evaluation criteria, new funding models, and new approaches to verification will be better positioned than those that wait for the full impact to become undeniable."

The window to act is narrow. These findings capture the impact of earlier AI models—the tools are only getting more powerful, and researchers are only getting more creative about how to use them.