How AI is transforming research: More papers, less quality, and a strained review system

The Flood of AI-Assisted Papers

A growing body of commentary out of academic circles, including recent reporting from UC Berkeley Haas News, is spotlighting a shift that many researchers have quietly noticed for the past two years: generative AI tools are dramatically accelerating how quickly scientific papers get written. Mathijs De Vaan, an associate professor of management at Haas, frames the issue through a familiar academic pain point — that writing up research has always been as laborious as the research itself. Large language models now compress that labor, letting scholars draft, revise, and polish manuscripts in a fraction of the time it once took.

Why Speed Isn't Free

The obvious upside is throughput — more findings reach preprint servers and journals faster than ever. But the tradeoff, as De Vaan and others describe it, is a subtle erosion of quality control. When drafting friction disappears, so does one of the natural checkpoints that used to force researchers to think harder about whether an idea was fully baked before submitting it. The result is a swelling volume of submissions that strains an already overworked peer-review system, where qualified reviewers are volunteers stretched thin across an expanding pile of manuscripts.

This matters well beyond management science. In fields adjacent to AI itself — LLM reasoning research, benchmark evaluations, and machine learning papers posted to arXiv — the same dynamic is arguably more acute. AI researchers are both the producers and heavy users of these writing tools, and the arXiv preprint ecosystem has already seen submission counts climb sharply in categories tied to machine learning and natural language processing. A system built on the assumption that most submissions represent months of deliberate effort is now absorbing papers assembled with AI assistance in days.

The Ripple Effects for AI Benchmarking

The concern is particularly pointed for benchmark-driven subfields, where claims of state-of-the-art performance or novel reasoning capabilities depend on careful experimental controls. If review capacity can't keep pace with submission volume, weaker papers — including those with inflated benchmark claims or insufficiently scrutinized methodology — may slip through or linger in an ever-growing preprint backlog. That has downstream consequences for anyone tracking the field: journalists, engineers, and other researchers building on reported results may find it harder to distinguish rigorously validated findings from AI-polished but under-tested ones.

What Comes Next

None of this suggests AI is uniformly degrading research quality; many scholars use these tools responsibly to clarify writing rather than shortcut substance. But the structural strain on peer review — a system already reliant on unpaid, time-constrained volunteers — appears to be a genuine and growing challenge. Expect more discussion in coming months around reviewer incentives, AI-assisted review tools, and whether journals and preprint servers need new triage mechanisms to keep pace with a faster, noisier publishing pipeline.