Why does science's publishing system seem to reward familiar work over the new ideas that change a field?

This explores why novel, field-changing research might struggle to get through peer review and gain recognition. A caveat first: the collection doesn't have papers that directly measure how disruptive a piece of work is, or that track how rejection rates differ for unconventional work. What it does have is evidence about the machinery of publishing. Taken together, that evidence suggests the system increasingly rewards work that resembles what came before.

The clearest signal comes from a large-scale study of AI-augmented researchers. They publish about three times as many papers and collect nearly five times as many citations. Yet science as a whole covers fewer topics and has less collaboration between researchers, because AI pulls effort toward problems that already have plenty of data Does AI help individual scientists while narrowing scientific focus?. This is the uncomfortable part: the tools that make individual careers more successful also pull the field toward well-mapped ground. Disruptive work usually starts where data is thin, so it competes for attention against a growing flood of productive, well-cited work on established questions.

The second force is evaluation itself. One study fine-tuned models on publication records and found they could predict which research pitches would land in top journals better than expert reviewers could Can institutional publication records train better scientific evaluators?. That sounds like progress until you notice what the models learned: they picked up each field's existing pecking order, not any written standard of quality. An evaluator trained on past success is very good at spotting work that looks like past success. As AI moves into reviewing, the risk is that the conservative instincts of human gatekeepers get automated and scaled. Agentic reviewers that check proofs line by line can catch real errors that humans miss Can inference scaling help reviewers catch errors humans miss?. That kind of rigor helps, but it checks correctness, not whether an idea opens new ground.

Third, the review system is overloaded and under attack. A survey of 230 publications describes a coupled arms race: AI scales up paper production, reviewers automate in response, authors manipulate the automated reviewers, and defenses and evasions follow Does AI create a coupled arms race in research production and review?. Researchers have found hidden prompts in arXiv manuscripts telling AI reviewers to be generous Are hidden AI prompts in preprints a deceptive research practice?. Fraud runs through organized paper mills and cooperating editors Does scientific fraud operate through organized networks or individual actors?. Fully AI-generated papers can clear workshop review Can AI-generated papers pass peer review undetected?. Conferences respond with detectors and stricter checks How can conferences detect and handle LLM misuse in peer review?. When gatekeepers are defending against noise and fraud, unusual work can look suspicious. The papers here don't measure that cost, but it follows from the dynamics they describe.

The surprising flip side is that recognition is no longer controlled only by peer review. An unreviewed preprint can shape a whole debate before anyone checks it, as MIT discovered when it disowned a widely discussed AI-and-science paper too late to undo its influence Can unreviewed preprints shape scientific debate before peer review?. So the bottleneck for new ideas may be shifting from 'can it get published?' to 'can it be heard over the volume?' In that contest, a careful, unconventional paper has no built-in advantage over a fast, confident one.