Can AI-generated research outpace peer review systems?

Nature's editorial, published alongside Sakana AI's peer-reviewed account of The AI Scientist, argues that AI-generated science has crossed from preprint curiosity into a published artifact, and that the pace of this change means "universities, funders, publishers and researchers must plan how they will adapt." The system "was able to follow this process and generate a research paper about its (negative) result," and the paper "passed the first round of peer review for submissions to a workshop" at ICLR, though it "did not meet the bar for the main conference track." The editorial also points to a GPT-5–assisted theoretical-physics preprint that physicist Nathaniel Craig called "journal-level research," and notes that Google, OpenAI and Anthropic are all now "trialling ways to automate research."

The reasoning is about scale and incentive, not just capability. Cheap, fast paper generation risks "automated, large-scale P hacking" that could "overload conference, publishing and funding peer-review systems... without shifting the needle on discovery." It is "hard to trace a model's inspirations," which risks using others' ideas without credit, and AI output "shreds the long-standing (if rough) correlation between authors applying effort and the work having value" — leaving open how hiring and promotion should treat AI-inflated output, and what happens to early-career researchers if tasks "crucial to their training" are done by a machine instead. Nature's own response is concrete: it "requires transparency in how LLMs are used," "will not accept such models as authors," and "encourages researchers to submit transcripts of prompts and model responses alongside the final outputs" for reproducibility.

The editorial cites, as its own evidence for narrowing diversity, the same finding detailed in Does AI help individual scientists while narrowing scientific focus? — that AI adoption make researchers more productive but shrinks the range of topics studied. It also shares the worry in Will AI automation widen science's productivity versus progress gap? that more output does not mean more discovery, framing the risk as review-system overload rather than a metrics-gaming incentive. Where Can automated review scale AI paper evaluation reliably? treats automated review as the fix that makes scale manageable, this editorial treats the resulting scale itself as the open problem institutions haven't solved. The specific paper behind the editorial sits in the same lineage as Can AI systems generate research papers that pass peer review?, an earlier version of the same system rather than the same study.

As an editorial, the piece asserts urgency and states Nature's own policy response; it does not measure how much AI-generated output already exists, how fast it is growing, or whether peer-review systems are in fact being overloaded today. Its strongest supporting examples — one workshop-accepted paper and one physicist's offhand "journal-level" assessment of a different preprint — are anecdotes, not a survey of the field. The defensible reading is that a major journal is treating institutional adaptation (authorship rules, credit, transparency, training continuity) as urgent and worth acting on now, not that the problems it names have already overwhelmed the system.