New research has significantly improved our understanding of the costs of the grant system with a detailed study by Alex Pollitt, Clare Taylor, Niall Sreenan, and Jonathan Grant, published in April 2025. Drawing on a number of grant calls and a large survey, the report shows that the cost of the grants system is around 13% of the funding allocated, with almost 90% of those costs arising from the preparation of applications. Importantly, the costs of peer review and administrative processes within funders are a low proportion of the overall cost. While consistent with previous analyses, the paper is the most thorough to date.

The key finding is the concentration of the costs on application preparation. Not only do these fall on researchers, the bulk of them relate to the preparation of unsuccessful applications because success rates are generally low in most research funding systems. The schemes that were studied had success rates between 19% and 31%. Because of the distribution, the overall costs of the process are strongly dependent on success rates, with expenditure rising as success rates fall.

This dependency on success rates means the cost of the system is highly sensitive to the balance between supply, the amount of funding available, and demand, the number of researchers seeking funding. In many systems, this relationship is out of balance, with the potential for escalating costs relative to the funding distributed.

While this cost issue is one threat to the grants system, there is another emerging challenge: a massive increase in the number of grant proposals driven by the use of AI to support their preparation. This threat has recently been described and documented by Geraint Rees and James Wilsdon. They report an increase in funding application numbers that coincides with the widespread availability of Large Language Models, a trend likely to accelerate further as agentic AI systems become more widely used. Higher volumes are not accompanied by lower quality, with many funders reporting that distinguishing between proposals is becoming increasingly difficult.

It might be comforting to think that at least the use of AI in the preparation of proposals will be reducing costs, but I think that comfort is misplaced. While the cost per proposal may be falling relative to those measured by Alex Pollitt and colleagues through surveys conducted in 2022 and 2023, this is being compensated for by an increasing number of proposals. Costs are at best static, while the number of unsuccessful proposals continues to grow.

Taken together, rising volumes, falling success rates, and increasing costs pose an existential threat to the grants system. Funders and the peer review system will struggle to cope with the volume of applications. Ever-increasing researcher effort will be absorbed by producing proposals, even with the help of AI.

Funders are struggling to respond. For example, increasing application volumes have recently led the European Research Council to propose and then soften new approaches to proposal resubmission, a form of demand management.

The grant system needs to change and incentives don’t necessarily help. Researchers want funding (as do their employers) and grant success can be career-defining. They won’t be supportive of solutions that reduce the volume or regularity of funding. Funders, on the other hand, get their reputations and right to exist from running rigorous and effective selection mechanisms. They won’t be supportive of changes that remove their power and influence.

Nevertheless, there are lots of ideas being discussed, which essentially fall into two categories: tweaking the existing system or radical redesigns.

Front and centre in the ‘tweaking’ category is demand management, where funders introduce rules designed to limit the number of applications that enter the grants system to reset the balance between supply and demand. These approaches tend to be unpopular with researchers, and in many cases don’t address the workload issue. Proposals still need to be written, but are filtered at the level of research organisations. Funders don’t see as many proposals but the work of producing them still goes on.

Perhaps a better approach to demand management is to place limits on how many proposals each researcher can submit in a given period. This could be done either with a cap on numbers (e.g. one proposal per year) or by a ‘pre-lottery’ where there is a random allocation of the right to submit proposals. The latter could be more sophisticated than a simple draw, with chances of success being weighted by how recently a proposal has been submitted or career stage. Either way, caps or pre-lotteries would need careful design, to avoid providing either too strong or too weak incentives for collaboration. In the face of constraints on application numbers, AI is likely to be redirected to maximising the quality of applications rather than the volume, but this will make the task of sorting proposals ever more challenging.

Another proposed solution is the use of simpler application processes, or two-stage processes where a short expression of interest is required with only a proportion of those being selected for a full proposal. I am not sure this has been examined rigorously, but anecdotal evidence suggests that short outline proposals generate many more applications; although the effort per proposal might go down, the total effort doesn’t. This phenomenon will likely be exacerbated by the use of AI in proposal preparation.

Changes to the review process have also been proposed. While partial randomisation can be an effective choice when faced with a high volume of applications, the need to produce a proposal in the first place means this change has limited impact on either volume or cost. Distributed peer review, where those who make applications are required to conduct the peer review, does have some potential. If researchers know that making a grant submission involves a reviewing commitment, then that may temper the number of applications that they submit. The use of AI for reviewing could undermine this potentially positive intervention.

These solutions focus on modifying the existing processes, but there are also proposals that advocate for wholly new methods to allocate funding.

One radical proposal is a universal research income, where every researcher gets a fixed annual allocation of funding without any application process. While superficially attractive, this approach has several problems. First, the allocations of funding are going to be too small to be meaningful. In a system where demand is outstripping supply, universal income is going to spread resources very thinly. Second, there is the difficult question of how to account for disciplinary, or even sub-disciplinary, differences in research costs. Finally, the use of a basic income for research removes governments’ ability to have strategic priorities for research. The system might work for investigator-led research but not for strategic or directed research programmes.

A variant of the basic income approach is to increase the existing organisational block grant funding, with the explicit requirement that the funding is used not only to support the underpinning of an effective research system but also to fund the conduct of research itself. This option is worth considering, although it would shift some burden of distributing funding to institutions. Like other options, it is also more suited to supporting investigator-led rather than more directed research.

The final radical approach is a switch to a more people-focussed selection process, a variant on the basic income approach where only a subset of researchers are provided with funding on the basis of a periodic individual assessment. Of course, there are already fellowship schemes, but the idea would be to allocate a much higher proportion of funding this way with the elimination of project-based proposals.

The key issue for this option is how the selection of researchers is made in a fair and transparent way, but at least the scale of the process (while large in a system like the UK) is not susceptible to AI-driven increases. Application processes could include interviews, focusing on individual knowledge and ideas.

There are some downsides to this individual-based approach. It would inevitably create a two-tier system, with some academics locked out of research funding, at least for a period. Or maybe permanently, as it would be hard to demonstrate a research track record following a period of no funding. The financial costs of people-focused selection are not known, although there is something to be learned from existing fellowship schemes. Finally, this approach could create issues for strategic research. If a significant proportion of academics aligned to a strategic theme were unfunded, it would limit potential progress in a priority area.

The grant funding system needs to change. While there have been critiques of the system for many years, the evidence suggests that abrupt and disruptive change is happening. All of the options under consideration merit further work, careful review and experimental testing where possible. My view is that a potential solution lies in a hybrid approach. For investigator-led research, I would favour application-level demand management, either a cap on the number of proposals per researcher in a given period, or some type of pre-lottery process. This approach regulates the volume of applications without creating additional administrative pressures for research organisations. And while it wouldn’t be popular with researchers, it is at least fair. For more directed research calls, distributed peer review offers the potential to impose a limit on the number of applications but this does need to be tested.

My proposed approach doesn’t touch on the more radical options for change. Of these options, perhaps the most promising is the expansion of fellowship schemes focussed on individual track record. It would be sensible to scale up fellowship approaches in parallel with the reforms outlined above. This would allow for a fuller evaluation of the costs and implications of a wholesale shift, and time to explore mitigations for the downsides

Time is short for managed reform of the research funding system. Funders need to act quickly, potentially in the face of incomplete evidence, while incremental adjustments are still possible.

The CBP report was produced in response to the UK Government’s AI for Science Strategy, which proposed a ‘missions’ approach to use of AI, launching one mission, accelerating drug discovery, and committing to further missions. The CBP proposed three new missions: pathogen detection, environmental forecasting and metascience.

The proposed focus for the AI for metascience mission is on the development of tools to improve the quality and speed of the processes of research, such as automated literature reviews, enhanced peer review and rigour assessment, and detection of errors. The authors argue that this has the potential to improve the quality of UK research and propose a target of increasing the proportion of research assessed as world-leading from 20% in REF 2021 to 30% in REF 2029. The potential benefits are clear, even if the target is poorly chosen. REF 2029 will assess research published over the eight years from 2021 to 2028.1 Five years of that period have already happened so around 63% of the eligible research outputs have already been published. And of the remaining 37% many will have already been submitted for peer review. It is too late in the REF cycle for a policy intervention to significantly alter the outcomes, and certainly to bring about an increase in research quality on the scale proposed.

Setting the target aside, the proposed interventions are indeed valuable. Using AI-enabled tools to assist in the process of research will bring about improved quality and impact. But they don’t necessarily help with making sure that the right research is being done, or that the allocation of funding is optimal. Careful and thorough metascientific studies can answer questions like ‘what funding structures correlate with higher-impact research?’ or ‘do methodological weaknesses cluster by field?’. Achieving these aims needs better insights into what works in research policy and funding, and there is huge potential for AI to drive progress in this aspect of the field. This is not without risk, as it will be important to continue to critically review the findings from AI-enabled research, being careful to consider biases and blind spots. Whether metascience is being conducted by researchers alone or augmented by AI, there remains scope for misleading conclusions.

The emerging power of agentic AI tools to write code and manipulate large datasets is becoming well known, and these tools have a general application in all research areas. Metascience is, however, particularly well-suited to the use of these tools. For many questions in metascience there are already large datasets available, both in the public domain and commercially available. The principal barrier to using these data to address new metascience questions is writing the code to manipulate, link together, and analyse existing data, and this barrier can be significantly addressed using agentic systems.

Tools are already emerging that demonstrate this. Dashun Wang and colleagues have produced a prototype—SciSciGPT—that is specially trained and tuned to answer questions in metascience provided as natural language queries. This system is based on a limited dataset (it only includes papers with US authors), but demonstrates the potential. The SciSciGPT system runs all the analysis in the cloud, is simple to operate, and has been pre-trained to understand the world of metascience research.

Alongside bespoke tools such as SciSciGPT, generalist agentic tools like Claude Code can enable metascience research to be conducted. My personal experience with Claude Code is that someone with limited technical expertise can conduct genuine metascience research. Merging and linking datasets, writing analytical code, and generating visualisations of the results are all within easy grasp, bringing the ability to conduct these types of analyses to a much wider range of people. For example, research funders or strategic leaders in research organisations no longer need deep technical skills to carry out metascience.

The CBP report proposes a role for AI in improving the pace and quality of science. What we need is better strategic intelligence to drive research forward, and AI has a major role to play there. AI will bring the tools of metascience into the hands of those who make decisions about research, closing the loop between intelligence and action. That’s a mission I can get behind.

I haven’t posted there for five years, and I want to pick up the previous themes—societal impact, metascience, research policy broadly—in a new format on Substack. One of the major developments that I will also cover is the ever-increasing impact of Artificial Intelligence, though, despite the em-dashes in the last sentence, AI won’t be doing the writing here. The writing will continue to be my personal take, although I may pick up on themes emerging from my new role at Digital Science. I am aiming at one or two posts per month, so the newsletter won’t take over your inbox. If you want to hear what I am thinking, please do subscribe.

Subscribe now