AI exposes what grant review really measures
Grant application overload, and funders’ attempts to address it, make a focus on the research that doesn’t move forward all the more important
Consider a researcher who wants to know how AI-driven news systems serve a language spoken by a small population. There is no established dataset, no agreed measure of success and no ready partnership with a news organisation – the first year of the project might have to go into creating the conditions for research.
Now consider a second researcher proposing to study a widely used English-language system, with data, measures and preliminary results already in hand. A panel could reasonably call the second project more feasible. But if readiness repeatedly decides the competition, the first language stays poorly studied because it was poorly studied before.
This difficulty is familiar to us. One of us directs a research centre in Norway whose work on news and recommendation systems requires infrastructure that English-language research can often take for granted. A good question can be clear long before the means to investigate it exist. Someone has to pay for the work between those two points.
Funders and universities need a way to support worthwhile questions that cannot yet produce a competitive application. We propose limited preparation grants, judged by the uncertainty they resolve, and university support that is accountable for the research it makes possible. And the wider debate over AI-written applications helps explain why this is needed.
In Nature, Geraint Rees and James Wilsdon have recently proposed shifting the emphasis of evaluation towards investigators, their teams, and their previous and ongoing research programmes. They explicitly recognise the risk of cumulative advantage and suggest protections including ring-fenced funding. Their proposal makes sense as a response to applications that are increasingly easy to optimise. But it also raises a question that safeguards for particular applicant groups cannot settle on their own: how does a researcher build the record that evaluation will reward?
Defensible decisions
As polished prose becomes less informative due to the widespread use of AI in grant applications, panels may place greater weight on preliminary data, established collaborations and evidence of delivery. These can be valuable indicators of whether a project will succeed – but they also reflect earlier opportunities to do research. A system that relies more heavily on them risks turning the consequences of previous funding decisions into reasons to repeat those decisions. AI makes it harder to avoid asking which parts of apparent merit are the products of prior investment.
The evidence for cumulative advantage predates these tools. In a 2018 study of Dutch research funding, Thijs Bol, Mathijs de Vaan and Arnout van de Rijt compared early-career applicants on either side of a funding threshold with nearly identical review scores. Over the following eight years, those who narrowly won accumulated more than twice as much subsequent funding as those who narrowly lost; part of the gap arose because unsuccessful applicants were less likely to apply again. The study covers one national system, and did not find that the later funding advantage was explained by research achievements enabled by the first grant. Earlier funding itself appeared to be an asset.
Experience nevertheless matters. One of us has previously argued for giving documented delivery more weight than application-writing skill. Researchers who write convincing proposals and researchers who can carry out the work are not always the same people. The difficulty is distinguishing between what a team has learned and the advantages that enabled it to learn. A young researcher can demonstrate competence through a study, a dataset or a piece of software, but producing each requires time and resources. Adjusting expectations for career stage helps; it does not supply those resources.
Funders therefore have two responsibilities. They must establish whether an applicant understands a project and can carry it out, but they must also consider whether their portfolio expands the range of questions that can be investigated. A sequence of defensible decisions about individual applications can still leave an important area without the capacity to compete. An individual review panel cannot remedy that outcome on its own, which makes institutional responsibility especially important.
Concentrating opportunities
The philosopher Philip Kitcher's account of well-ordered science, developed in Science, Truth, and Democracy, connects research priorities to informed deliberation among people who take one another's interests seriously. Expert assessment can establish whether a method is credible or a question has already been answered – it cannot by itself establish that the current distribution of research attention is worth preserving. Helen Longino's The Fate of Knowledge adds an epistemic reason to care: inquiry depends on criticism from people able to challenge its assumptions. Concentrating opportunities can restrict that criticism.
Participation in setting priorities need not weaken standards of evidence. The James Lind Alliance brings patients, carers and clinicians together to identify and prioritise unanswered questions about health. They help determine what deserves investigation; the resulting studies still have to establish their findings. Similar attention to whose questions remain unasked would help funders distinguish a lack of research readiness from a lack of importance. Basic science and scholarship must retain room for questions whose value is not yet evident.
Universities can begin this process before a question reaches a funding agency. Research development staff often encounter ideas that are set aside because there is no dataset, no accessible community or no suitable partner. Under pressure to increase grant income, institutions have a reason to concentrate support on applications most likely to succeed. That is understandable – but it can also leave promising ideas without a route to the point at which their merits can be tested.
Institutions already invest in preparing applications. Norway's Research Council, for example, offered up to NOK138 million for 2025–27 to help institutions develop Horizon Europe proposals. Such support can improve project design and coordination as well as presentation. Given that public funders already accept a role in paying for preparation, the next question is how much support should reach ideas whose main obstacle is the absence of research infrastructure, especially when no suitable competition is immediately available.
Universities could reserve a small share of discretionary research development funding for those cases. Applicants would explain why a question matters, what prevents investigation, and what a modest award would allow them to learn. Support might pay for testing a measure or negotiating access to data; success would include finding that an approach will not work. An office's performance should reflect what its support made possible, alongside the awards it helped secure. This would give staff a reason to develop ideas whose value exceeds their immediate prospects of winning a grant.
The ideas that don’t go forward
AI is making the question of who pays for preparation more urgent. In September, UK Research and Innovation (UKRI) announced changes to grant assessment for the age of AI, citing very high application volumes and low success rates for some funding opportunities.
Its changes include more active demand management, interviews for higher-risk awards, and a retrospective trial comparing what AI would have concluded about past applications with the decisions actually made. UKRI says expert human judgement will remain at the heart of every decision. These measures address real pressures on review, and interviews can check whether applicants understand what they propose, which polished prose now shows less reliably. But each acts on applications that researchers were already in a position to prepare, and none helps a question become ready for review.
Demand management also moves part of the selection process into universities. Where a funding opportunity caps the number of applications each organisation can submit, universities must decide which go forward and will tend to pick those most likely to succeed. Institutions should then also record why the other applications stop and what support might change that decision. Otherwise, reductions in application volume could conceal the repeated exclusion of questions that need preparation. A short account of the barrier would be more useful than another ranking, provided that institutions use it to guide support and avoid creating a new reporting exercise for every unsuccessful idea.
Funders can examine the same problem across their portfolios. Where expressions of interest are collected, agencies could follow broad topics through applications to awards and ask researchers what prevented particular ideas from progressing. They would need additional conversations to learn about those ideas which were never formally submitted. These records would be incomplete, but could reveal recurring obstacles. If a line of work repeatedly fails for lacking preliminary data, the agency should ask whether it offers a credible route to generating those data.
Preparation grants could provide that route. Exploratory funding already provides some of the answer: the NIH's R21 mechanism, for example, generally does not require preliminary data unless the funding opportunity specifies otherwise. The test is whether existing provision reaches the particular questions and researchers held back by missing infrastructure; where it does not, a limited programme could fund the work needed to establish feasibility, with criteria suited to that purpose.
An application should identify a specific uncertainty and a proportionate way to resolve it. A team seeking access to a community, for instance, would explain what relationship it needs to establish, why that relationship matters to the research, and how it would learn whether collaboration is feasible. It should not have to present an established partnership as proof that it deserves support to build one. Assessment would still examine the value of the question, relevant competence and the proposed use of funds. Applications should be short, and reviewers should assess competence for the preparatory task without requiring an established funding record.
Recipients would report what they learned and, where appropriate, make resulting materials available for further research. Evaluation should ask whether the grants produced usable resources, enabled previously impractical studies and reached researchers outside established funding networks. Subsequent grant income would be one possible outcome, rather than the sole measure of success. A well-supported finding that a method is unsuitable or access cannot be secured can prevent a much larger waste of money later.
Reconsideration
There are costs. Money spent on preparation cannot also fund a team ready to start tomorrow, and some preparation will lead nowhere. Universities and funders should make that trade-off explicit, begin with a limited allocation and change or stop programmes that fail to expand research opportunities. But the alternative also has a cost: valuable questions remain persistently infeasible because the work that would make them feasible is nobody's responsibility.
AI gives funders a reason to reconsider what applications can tell them – and that reconsideration should extend to the opportunities that make an applicant credible in the first place. Reviewers work within programmes whose aims and budgets they did not set. The institutions that design those programmes must decide which uncertainties they will pay to resolve.
A funding system should be able to explain how it helps important questions become researchable, as well as how it chooses among those which are already ready for review.
Christoph Trattner is a professor of information science at the University of Bergen, Norway, and director of SFI MediaFutures. Emily LaRosa is an AI philosopher and researcher at the University of Bergen.