Does a paper's citation count say more about whether its science is good than the prestige of the journal that printed it?

This explores whether citation counts or venue prestige better tell you a paper is good. The short answer: the collection doesn't settle this directly. What it shows is more interesting. Each signal holds real information that models can learn from, and each can be distorted in its own way.

Start with the case for each. One group fine-tuned models on where social science research pitches ended up being published. Those models beat both frontier reasoning models and majorities of expert reviewers at predicting outcomes, reaching 59.2% accuracy in management, where experts agreed only 41.6% of the time Can institutional publication records train better scientific evaluators?. The models learned a field's unwritten standards just from which journals accepted what. A separate project took the citation route. It trained models on 700,000 pairs of papers matched on citations, and they learned a kind of 'scientific taste': they predicted impact better than GPT-5.2 and came up with higher-impact ideas Can models learn what makes research worth doing?. So venue tiers capture what gatekeepers value, and citations capture what the community later uses. Those are two different things, and both can be learned.

Now the cracks. Citations can grow for reasons that have little to do with quality. Researchers who use AI publish about 3× more papers and get 4.8× more citations. Over the same period, science as a whole covers fewer topics and has less collaboration, because work piles up on data-rich problems Does AI help individual scientists while narrowing scientific focus?. A high citation count can reward following the crowd. A finding from a different setting suggests why counts carry weight anyway: users prefer AI answers with more citations almost as much when the citations are irrelevant Do users trust citations more when there are simply more of them?. People treat volume as credibility.

Venue tiers have their own weak spot: the peer review that produces them. One model shows how a flood of submissions overloads unpaid reviewers. Review accuracy drops, authors respond by submitting more speculative work, and submissions climb further Does peer review quality collapse under submission overload?. A survey of 230 publications describes an arms race. AI scales up paper production, AI-assisted review tries to keep pace, and manipulation and countermeasures escalate together Does AI create a coupled arms race in research production and review?. Conferences are now desk-rejecting papers with made-up references How can conferences detect and handle LLM misuse in peer review?. Meanwhile, some work goes around venues entirely. One arXiv preprint shaped public debate long before MIT said it had no confidence in it Can unreviewed preprints shape scientific debate before peer review?.

The surprising takeaway is that both signals are now being turned into AI evaluators. Whatever bias a signal has, the evaluator trained on it will inherit. A model trained on citations may learn to favor crowded, data-rich topics, and a model trained on venues may learn to favor whatever an overloaded review system lets through. The better question may be which kind of distortion you're willing to automate.