Science has never had more ways to produce knowledge. There are more researchers, papers, datasets, collaborations and computational tools than before, and artificial intelligence promises to speed up the work again. But a larger stream of research does not necessarily mean a wider search for answers. A scientific system can generate more papers while testing fewer genuinely different possibilities.
That is the paradox at the center of Nicola Cinardi’s Perspective, “The directions of science: How we search the unknown.” The paper asks us to look beyond the volume of research and examine how science chooses what gets pursued, supported, produced and remembered. Its concern is not that science should always chase novelty, or that every neglected idea deserves funding. It is that when the future value of competing directions is unclear, repeated selection through overlapping signals may cause a field to settle too early.
The question matters well beyond universities. Scientific research helps societies respond to disease, climate change, food insecurity and technological risk. Those problems rarely arrive with a label identifying the best method or theory in advance. We need systems capable of concentrating effort when evidence is strong—and of keeping alternatives alive long enough to learn when it is not.
The Science
Cinardi describes science as a collective search through a changing landscape of possible research trajectories. A trajectory is more than a topic: it can combine a question, a hypothesis, a method, a way of representing a problem and a particular empirical setting. The landscape changes as discoveries create new questions, instruments make new observations possible, and methods migrate between fields.
Because no scientific community can pursue every possible experiment or idea, selection is unavoidable. The paper groups the process into four functional stages. Entry concerns who enters, remains in and advances through scientific careers. Allocation concerns who and what receive money, positions, infrastructure and time. Production is the work of turning possibilities into experiments, analyses and outputs. Retention concerns which findings persist through publication, citations, teaching, replication and future research.
These are not four separate boxes in an organizational chart. A university may shape careers and provide resources; a journal may influence both publication and what researchers choose to study; a software platform may help generate hypotheses and process results. Cinardi calls any mechanism that changes the odds that a research possibility gets explored or survives a “selector.”
The paper’s main concept is cross-stage dependence: selection at one stage influences later selection, or different stages rely on overlapping criteria, information or infrastructure. Imagine two plausible directions, A and B, with no decisive evidence between them. A receives a grant, enabling more experiments. Its papers become more visible, and its prominence helps it win further grants. Later researchers encounter A more often and build on it. The result may be that several institutions appear to have independently concluded that A is best, when some are responding to a signal created by the first award.
That does not prove A is a bad choice. It might truly be the stronger direction. The issue is whether the system has generated enough independent evidence about B to know. Agreement among funders, journals and researchers can reflect reliable shared evidence; agreement alone is not proof of dependence. The challenge is to distinguish repeated reactions to an inherited signal from separate judgments based on new evidence.
The paper builds on research about the division of scientific labour, collective search and institutional incentives. Earlier work has explored how researchers spread effort across competing approaches and how communication structures influence which hypotheses survive. Studies have also documented cumulative advantage, where an early success improves later opportunities, and path dependence, where past choices make some future routes easier than others. Cinardi’s proposed focus is narrower: whether selectors at different functional stages offer distinct grounds for choosing which scientific directions continue.
What They Found
This is a Perspective, not a study that estimates how much cross-stage dependence exists across science. It reports no new dataset, numerical effect size or global measure. The contribution is a conceptual framework, a conditional prediction and suggestions for how researchers might test them. That distinction is essential: the paper raises a plausible risk; it does not establish that science as a whole is already trapped in reinforcing loops.
The evidence it discusses shows that potential links exist. Early funding success can improve later funding chances, including among applicants whose initial evaluations were very close (Bol et al., 2018). Yet the effects are not inevitable: research on NIH grants found that receiving an R01 grant produced only a modest increase in subsequent publication output, consistent with unsuccessful applicants finding other support (Jacob and Lefgren, 2011). Other work suggests that the same research can be evaluated differently depending on how prominent its author is presented, and that evaluators’ intellectual distance from a proposal can affect scores (Huber et al., 2022; Boudreau et al., 2016).
These findings do not add up to a single verdict. They point to mechanisms that can transmit earlier selections, as well as conditions that may interrupt transmission. The Perspective also notes that research organization and incentives shape what is produced: smaller and larger teams have been associated with different roles in innovation, while evaluation systems can reward work that is easier to assess or appears more productive. Publication and attention then help decide which outputs become the starting points for future work.
The author’s central prediction is conditional. When uncertainty is substantial, and systems have comparable visible diversity, greater cross-stage dependence should make it harder to explore alternatives and adapt when new evidence changes their relative promise. In practical terms, alternatives may lose support more quickly, and a field may redirect effort more slowly after a favoured route disappoints. If high dependence instead improves adaptive exploration—or if it has no relationship to these outcomes—the proposed prediction would not be supported.
That is why the paper does not offer charts of measured effects: its central quantity has not yet been measured here. Turning the argument into an empirical research programme would require experiments, natural experiments or longitudinal comparisons. Researchers could examine, for example, whether a grant awarded through a close-call decision affects later funding or attention beyond the scientific results it enabled. They could compare evaluations of otherwise identical work when prior funding, author prestige or citation information is visible versus hidden.
The paper’s four-stage map—entry, allocation, production and retention—is a useful way to locate where selection happens. But the important feature is the links between stages, not the boxes themselves: a decision may travel forward as funding, reputation, infrastructure or visibility.
The hypothetical A-versus-B example illustrates how a direction can gather momentum through successive choices. It is an explanatory scenario, not observed data, and should not be mistaken for proof that the favoured direction is scientifically inferior.
Why This Changes Things
The distinction between visible and effective diversity is the paper’s most useful shift in perspective. A field can contain many researchers, laboratories, funders and journals and still offer relatively few different routes through which ideas are evaluated. If those actors rely on the same publication metrics, prestige signals, datasets or AI systems, their numbers may exaggerate how many independent judgments the field really makes.
This matters because novelty and exploration are not ends in themselves. A medical research field should not maintain an ineffective treatment indefinitely simply to preserve diversity. A mature theory supported by strong evidence deserves concentrated effort. The problem appears earlier, when evidence remains incomplete and a plausible route can be eliminated because it has not yet accumulated the very signals that would make it look promising: publications, citations, prior funding or established expertise.
The cost of abandoning a direction depends on how reversible it is. A field can sometimes return to an underused question. But if researchers leave, specialized instruments are dismantled, datasets stop being maintained or institutional memory disappears, rebuilding that route may be expensive. Scientific options, in other words, can have an infrastructure and a workforce attached to them. Temporary neglect can become practical impossibility.
AI raises the stakes because it can contribute at multiple stages. Tools that search literature, suggest hypotheses, perform analysis and draft text can enlarge the set of things a research team can attempt. But if many researchers use similar models, training data, benchmarks or platform defaults, the tools could also make their searches more alike. The same systems may influence both the production of candidate ideas and the selection or presentation of results. This is a possibility the paper highlights, not an outcome it measures.
The public consequence is not simply “less innovation.” It is a question of how societies learn under uncertainty. If one direction is favoured early, it may attract more evidence and look increasingly convincing partly because it received the resources to produce that evidence. A system with independent ways to challenge its own early choices is better placed to discover errors, recognize a neglected opportunity or change course when circumstances shift.
That argument should not be mistaken for a call to scatter resources indiscriminately. Diversification has costs: money and attention spent on one route cannot be spent elsewhere, and maintaining every alternative is impossible. The relevant goal is not maximum diversity. It is enough independence and reversibility to make credible comparisons before the evidence settles the question.
What’s Next
Cinardi’s framework suggests concrete tests. Researchers can use funding lotteries, threshold rules or policy changes that create near-random differences in who receives support, then trace whether those differences affect later funding, hiring, publication or attention. They can also run controlled evaluations in which reviewers see or do not see signals such as prior grants, institutional affiliation or citation counts. Comparisons across institutions may reveal whether different selectors genuinely use different information—or share criteria and technical infrastructure.
Care will be needed. A funded project may later attract attention because it produced better results, not merely because it was funded. A credible test must separate those pathways. Conversely, if an early decision predicts later success only because it enabled valuable experiments, that is not necessarily a failure of independent judgment. The paper explicitly treats evidence against its hypothesis as informative, rather than assuming that concentration itself proves a problem.
The bigger question is how to design scientific systems that can both coordinate and correct. Shared standards help communities build cumulative knowledge. But where uncertainty is high, multiple ways to evaluate a question may help prevent an early signal from becoming an unquestioned verdict. For AI, that could mean asking not only how much faster tools make research, but whether they broaden the sources, methods and judgments brought to bear—or quietly standardize them.
Science does not need to pursue every possibility forever. It does need to preserve enough capacity to change direction. The measure of progress is not only how quickly a system produces answers, but whether it can still ask a different question when the evidence demands one.