Adjudicators to orchestrators: How public research funders are responding to AI

Why public research funders are struggling to scale AI

Like their stakeholders in government, academia and industry, public funders are increasingly looking to AI to deliver improved productivity, enhanced analytical capabilities and data-backed decision-making. Despite these aspirations, AI use and maturity vary considerably across the world. Even in the US, where federal agencies have identified more than 150 use cases relevant to research fundingopens in new tab/window, many initiatives are still at the pilot stage.

This is partly because of the relative youth of modern AI, with many of the most established projects having originated as text mining or traditional machine learning programs that relied on statistical models, with humans doing much of the heavy lifting. Another factor has been the need to grapple with now-familiar risks such as hallucinations, algorithmic bias and data privacy threats, while organisational challenges have also been slowing progress. Besides the perennial problem of cost, many funders face severe skills gaps and difficulties accessing quality data, often as a result of the very system fragmentation some of them hope to address with AI.

While some governments have well-defined AI policies, agencies can also suffer from a lack of clear guidance about how to best approach the technology. Typically funders operate within a framework of evolving legislative and regulatory requirements, often with no widely accepted way to measure the return on AI investments.

All of this can contribute to a culture of risk aversion – an approach one former US Chief AI Officer describes as “largely focused around preventing harm and making sure nothing bad happens”— where incremental changes are consistently preferred to more ambitious innovations.

Submission control, we have a problem

Why are grant applications surging?

This tension between caution and the desire for change is especially evident in funders’ responses to the boom in AI-optimized grant applications, with the sheer volume of submissions now threatening to overwhelm their systems.

To date, the best indicator of the scale of this problem is probably the ambitious survey of grant applications across 12 European funding bodies described in the 2026 Nature comment article Could agentic AI topple grant-funding systems?opens in new tab/window. The article notes that submission rates have risen by an average of 57% since 2022, when the original version of ChatGPT was released. This trend is not confined to Europe, with the US National Institutes of Health (NIH) reporting a similarly sharp rise in grant submissions, with early-stage investigator application pools rising by 11% in 2025 aloneopens in new tab/window. This has contributed to a dramatic drop in NIH grant success rates (to about 17%), a situation compounded by a decrease in the overall number of awards driven by government budgetary shifts.

While Asian research powerhouses like China and India have, by contrast, seen major increases in national science and technology spending, grant applications are still outpacing available funding. When the authors of the Nature article warn that “systems of grant funding and review will collapse unless funders adopt new strategies for managing volume and demand”, their words will clearly resonate differently in each of these countries, but the evidence of a growing AI-based submissions crisis is compelling.

While most funders agree that drastic action needs to be taken, there is little consensus on what the long-term solution should be. A number of agencies, such as the NIH and the Dutch Research Council (NWO), have introduced curbs on the substantial use of AI applications in proposal development, while the Chinese Ministry of Science and Technology has gone further and banned the direct use of generative AI in writing and applying for research funding.

Other funders, like UK Research and Innovation (UKRI) and the Japan Science and Technology Agency (JST), are more open to some AI involvement in applications (for example, for brainstorming or drafting) provided it is declared.

It is worth pointing out that these submission-control policies do not exist in isolation — they are typically accompanied by measures designed to prevent researchers from uploading proposals to non-secure AI tools or to prohibit the use of AI in grant review – but all these changes have been controversial and present significant enforcement challenges.

For example, while most funders are explicit about the need for human review, some, like the US Department of Health and Human Services (HHS), have been using AI-based software to detect the use of AI so they can triage compliant submissions to expert reviewers.

Commentators have noted that this AI-to-AI screening process threatens to obscure the genuine human credentials funders are looking for. It also highlights the contradictions implicit in their overarching approach to the new technology, where internal receptiveness is awkwardly juxtaposed with external caution. This conflicted environment — short-term pragmatism jostling with long-term planning, ecosystem conservatism competing against research revisionism — forms the backdrop to government funders’ growing engagement with AI.

How do government research funders currently engage with AI?

While government research funders are exploring a wide range of AI use cases, many of these can be consolidated into a few broad categories across, or adjacent to, the award cycle.

• Improving submissions – Before research funders became inundated by AI-assisted grant applications, many openly tolerated the use of the technology by researchers, suggesting it be used to help with accessibility, language barriers and specialized tasks such as drafting code. This thinking is still present in the guidelines of major funders like UKRI and the European Research Council (ERC), although the tone has understandably become more prescriptive and clearer declarations of AI use are routinely required from applicants. Interestingly, the overall standard of AI-supported grant submissions tends to be high – unlike AI-augmented job applications, many of which are highly generic and therefore relatively easy to dismiss. An overall increase in quality resulting from diligent AI-assisted searching, brainstorming and polishing arguably makes the selection process even more difficult.

• Screening and evaluation of funding proposals – Because this function is often considered the most resource-intensive aspect of the funding process (rejection ratesopens in new tab/window for major global government research funders typically hover between 80% and 90%), it has also been an early target for AI interventions designed to reduce administrative burden. Here, much of the pioneering work has been done by Mike Thelwall and researchers from the Universities of Sheffield and Wolverhampton (both UK), whose foundational papers indicate that large language models (LLMs) like ChatGPT can outperform traditional citation metricsopens in new tab/window in predicting human expert quality assessments. Also under the auspices of UKRI, the same team assessed whether it was reasonable to give large language models a role in the peer review of grant proposals. Their findingsopens in new tab/window showed that while AI scores correlate positively with expert review scores, they are ultimately still "educated guesses" and should not replace human gatekeepers. Indeed, most AI initiatives in proposal evaluation are focused on supporting, rather than replacing, human decision-making. For example, the La Caixa Foundation, a Spanish non-profit research funder, conducted a trialopens in new tab/window where an AI model screened out around one in six of its biomedical applications, with two human reviewers checking whether these rejections were fair before triaging the remaining proposals to the full panel of experts. Other initiatives have focused on using LLMs to check for similarity between proposals, or to benchmark the consistency of human-decision-making. AI tools also sometimes perform a tie-breaking function when reviewers disagree or are simply unavailable – a more informed alternative to the lotteries used by funders like UKRI to discriminate between top-tier applications deemed “equally good." Discussions around the longer-term role of AI in proposal evaluation are often highly emotive, with two distinct strains of thought emerging. The first of these can be described as conservative because, while the use of AI technology may be novel, its application tends to reinforce the existing structures of the research system. For example, one reason why the La Caixa Foundation trial proved so contentious is because the LLMs based their conclusions on matches between the language of submissions and that of previously successful proposals, tending to reinforce the status quo in a way that was not immediately transparent. The same thinking informs plans to shift the focus of evaluation away from written proposals and toward the wider context of the applicants: their previous funding successes, publication impact, collaborative reach, methodological diversity, career trajectory, research team and institutional prestige. While there is consensus in the research community around the need to broaden the definition of impact, there are also concerns that focusing on previous achievements will simply replicate the power relationships of the current system. This, the argument runs, compromises impartiality, limits novelty and neglects the unrealized “knowledge capital” of early career researchers. This brings us to the second strain of thought, which sees AI as a means of democratizing a research system that tends, in the words of one recent paperopens in new tab/window, to “privilege senior investigators while sidelining early career scientists and genuinely novel ideas.” The same author continues, “when bold ideas are consistently filtered out in favor of safe bets, the pace of discovery slows. At stake is society's ability to mobilize scientific creativity against its most urgent challenges.” While it is easy to get swept up in the call for “bold” (that is, disruptive) ideas to tackle societal grand challenges, we should be wary of simplistically conflating innovation with age or career stage, since there are significant variations across disciplines. It is, however, entirely plausible that AI can help counter the tendency of funding systems to favor established status, track records and networks, both in the case of individual researchers and of institutions that “lack the machinery of proposal production”opens in new tab/window that might be available to their wealthier peers. On a broader level, the technology might also help address the problems posed by language barriers or regional bias, as originally envisaged by UKRI, ERC and others. Finally, AI systems can promote innovative research projects by broadening literature scans, applying evaluation criteria with impartial consistency across thousands of proposals, or proactively detecting historically underfunded but later successful projects. There are multiple notes of caution to be sounded here – AI may remove some biases but introduce new ones of its own, uniformly applied AI criteria may not be effective against a large volume of AI-optimized proposals, while innovative researchers who become part of the system may stop being innovative because they are a part of the system – but nonetheless, the revolutionary potential is real.

• Streamlining peer-review and application processing – According to a major 2025 reportopens in new tab/window, identifying the best peer reviewers and panel members is currently the most common application of AI in research funding. The best-practice AI handbook, Funding by Algorithmopens in new tab/window, includes a case study from the Swiss National Science Foundation (SNSF), which sources prospective reviewers based on the titles and abstracts of their published outputs. While the Swiss program is supplemented by a final human-in-the-loop checking stage, any model based exclusively on past publication performance opens itself to charges that it reinforces the research system status quo discussed above.The US National Science Foundation (NSF) appears to take a more sophisticated approach to mining potential reviewers’ areas of expertiseopens in new tab/window, building expertise profiles, facilitating conflict-of-interest-aware matching and developing collaboration and network maps. Since 2024, its Technology, Innovation and Partnerships (TIP) division has conducted AI scans of potential reviewers' professional profiles on sites like LinkedIn, feeding the results into their internal databases.Given the failure of current pre-screening programs to manage the huge influx of AI-assisted proposal submissions, there has been discussion around how AI might augment the peer-review process itself – for instance, by supporting the production and processing of peer-review reports. While bodies such as the ERC have given their reviewers strict guidelinesopens in new tab/window on AI usage, this area is now being proactively explored. UKRI, for example, has begun investigating responsible applications of AIopens in new tab/window in the grant peer review process.

• Grant management – Government research funders have also begun to use AI to automate the monitoring of grant compliance. In the US, the Department of Health and Human Services (HHS) and the Department of Defense (DoD) are both piloting AI solutions in this area, while the cross-agency GovGrants platform, which incorporates AI features, includes compliance controls alongside AI monitoring for fraud, waste and abuse. HHS has also been exploring the use of AI to identify grants that may need to be reviewed for alignment with government executive orders that cover topics like Diversity, Equity and Inclusion (DEI) and gender theory.

• Impact tracking and identifying funded research – A recurrent challenge for funders is their inability to consistently track all the research they support, along with the specific grant(s) via which this is accomplished. This is strategically important because a comprehensive understanding of what and how projects have been funded should inform any agency’s impact reporting and strategic decision-making. The problem can partly be attributed to fragmented data holdings, incomplete project descriptions or “ghost research” that has not been made public (for example, for security reasons) and so does not appear in the normal records. While researchers are usually legally required to acknowledge grants in their published outputs, many do not comply – either because they are unaware of the requirements, or because the request gets lost in internal administration.Efforts in this area include the NIH Research, Condition, and Disease Categorization (RCDC) System, which employs the NIH Automated Indexing Service (NAIS) to match project titles and other data against a biomedical thesaurus of concepts. In another case study from Funding by Algorithmopens in new tab/window, the independent Novo Nordisk Foundation’s (NNF) AI-driven Acknowledgements Project has successfully leveraged text mining of article publication data and grant application data to match publications with their likely grants

• Impact tracking and real-world impact – Beyond tracking funded research, agencies are also beginning to use AI to understand the wider impact of their portfolios more effectively. While measuring the influence of research within academia using citation-based metrics is comparatively straightforward, although not without controversy, gauging how far outputs affect society, national economies, or health systems is especially problematic because of the long timescales involved and the difficulty of attributing broad changes to specific projects or articles. That said, government funding bodies must rise to this challenge if they are to be truly accountable to the taxpayers they serve.Again, text mining approaches predominate, with research assessments like the UK Research Excellence Framework (REF) training machine-learning models on case studies submitted by universities to identify impact topics. In 2021, the Research Council of Norway adopted a broadly similar approach, using AI-enhanced text mining to match the abstracts of funded articles with impact types like patents, social media citations and policy documents. Meanwhile, in the US, the Clinical and Translational Science Awards (CTSA) Program, part of the NIH, has adopted a more context-driven strategy, deploying natural language processing (NLP) to extract impact benefits from publications, patents and grants. The approach is guided by the Translational Science Benefits Model (TSBM), a structured taxonomy of Clinical, Community, Economic and Policy impacts.

• Portfolio analysis and strategic planning – Like universities and corporations, government funding bodies are also using AI to identify research trends to help target their investments. Typically, funders will ingest huge datasets of grant applications and publications and use predictive modelling to create programs that identify new, neglected or over-funded areas of research. For example, the 2024 versionopens in new tab/window of the European Commission Joint Research Centre (JRC) methodology for detecting early indicators of emerging technologies (“weak signals”) uses text mining, clustering techniques and scientometric indicators to peer-reviewed scientific publications and patent documents.Another common approach is to review the growth of collaboration patterns between disciplines, organizations and countries, surfacing topics or researchers that might be absent from a current grant portfolio. Once again, the aim is usually to select the best funding instruments (the mechanisms through which money is distributed) and call formulations (the texts that define what a given funding round is looking for) to meet specific research policy objectives. However, these techniques can also be used for internal reviews, for example the NIH Office of Portfolio Analysis (OPA) study which identified potential structural barriersopens in new tab/window facing institutions “outside high-centrality positions” that aimed to participate in multi-institutional AI research.

• Acceleration of scientific discovery – Like researchers, funders are keen to use AI to drive discovery and help manage the problem of information overload, frequently with an emphasis on expediting either literature reviews or systematic reviews. Efforts are often targeted at specific priority or “high return” areas – for example, UKRI has channelled investments toward AI-driven drug discovery, healthcare breakthroughs and environmental modelling, while the US Advanced Research Projects Agency–Energy (ARPA-E), part of the Department of Energy (DOE), is combining AI with self-driving laboratories to compress timelines for industrial catalyst development. The results of these set-piece AI projects can sometimes be spectacular, the most celebrated example of all being the AlphaFold AI programopens in new tab/window, which solved the 50 year-old protein folding problem, thereby helping transform medicine, agriculture and environmental science, by drawing on the data infrastructure of UKRI and other UK-based agencies like the Biotechnology and Biological Sciences Research Council (BBSRC), the Medical Research Council (MRC) and the Wellcome Trust. While the success of AlphaFold was built on decades of public investment and quality data infrastructure, most government research funders face significant obstacles in securing the external data sources they need to support AI discovery. There can sometimes be a cultural reluctance to use external resources. One former US Chief AI officer describes how his government contacts “were not very friendly towards the possibility of how AI can revolutionize healthcare by leveraging additional streams of data”² while another key problem is the licensing costs of proprietary datasets, often compounded by slow and bureaucratic procurement processes. In addition, there can be legal and privacy barriers, or difficulties ensuring that information from external providers meets internal data classification, handling and residency and sovereignty standards. Finally, experts within an agency might have an incomplete understanding of the broader data landscape. Well-resourced AI projects can fail simply because the right dataset is never identified.

• Supporting policymakers – Some agencies are also utilizing AI data-synthesis platforms to rapidly supply policymakers with reliable scientific information to inform their decision-making. The UK-based METIUS projectopens in new tab/window (Mobilising Evidence Through AI and User-informed Synthesis) specifically targets government staff who (like almost everyone else) are unable to keep up with the pace of scientific publication, attempting to surface the specific information they need in a readily digestible way. In the US, the ADVISE (AI-accelerated Design of Evidence Synthesis for Global Development) framework, which grew out of the evidence agenda of USAID, the international development agency, sets out to achieve similar goals, albeit with a narrower scope.

AI readiness for internal systems

Beyond these specific use cases, there are also a number of general issues that impact public funder’s ability to benefit from AI. While many agencies face difficulties securing the right external datasets for their in-house systems, they can also experience problems ensuring their own data is AI-ready.

This is problematic because funders often hold huge quantities of valuable information – funding applications, research assessment exercises, outputs and outcomes tracking, peer-review data, researcher and institutional career data, financial and expenditure data, collaboration and network data, policy and impact case studies – that is simply unavailable elsewhere. This data is often dispersed across different, unconnected systems, built by different IT contractors at different times, or even inherited through mergers like UKRI’s assimilation of seven research councils in 2018. Moreover, the data itself has not usually been designed for machine-based analysis and can vary widely in quality, with incomplete, inconsistent or overlapping records resulting from decades of organic change to policies and funding priorities. Although funders typically have considerable expertise in grant administration, many lack the in-house data science capacity that would help them to get a handle on this disparate mass of information. The fact that much of it includes personal details or otherwise sensitive data that dates from before the advent of widespread AI training adds a further layer of complexity, making funders understandably cautious about uploading it to their new systems.

There is, of course, an argument that AI can be used to help address some of the problems caused by fragmented data. This is an attractive idea because, while expensive and labor-intensive, the AI solution often seems less onerous than a full infrastructure overhaul. NASA, for instance, has developed the OCIO STI Concept Tagging Serviceopens in new tab/window, an AI-powered natural language processing tool designed to organize and automate keyword tagging for the agency’s diffuse collection of scientific documents and research papers. However, while AI is undoubtedly useful in structuring complex datasets, it is no substitute for an in-depth data strategy that includes centralized processes and properly integrated systems and policies. The same can be said of the broader trend for organizations to develop their own internal “walled garden” AI systems. Sovereign AI platforms may give funders scope to develop custom solutions, better management of bias and hallucinations and enhanced security, lessening their reliance on external vendors, but in the words of recent IBM researchopens in new tab/window, “without a unified environment that provides access to both structured and unstructured data, organizations will struggle to move AI projects into production at the speed and scale required to be competitive.” In other words, the old “garbage in, garbage out” axiom still holds true. Advanced AI trained on bad data generally means more sophisticated errors presented with a higher degree of confidence.

Is AI choking the research ecosystem?

Public funders are perhaps uniquely well qualified to bring the transformative benefits of AI to the research ecosystem because of their position at the crossroads of government, academia and industry, uniting policy, economic development and research. However, despite this influence and the innumerable projects exploring applications of the technology to their own workflows, we have seen that many of these organizations have been thrown into crisis by the unprecedented volume of AI-assisted grant applications. Funders across the world risk paralysis unless urgent steps are taken to control this problem, but the logistical threat also conceals a more fundamental challenge. These bodies are defined by the expertise that enables them to judiciously allocate public money to the research projects that will best serve society. Remove their capacity to make meaningful judgments – as the AI submission explosion threatens to do – and a number of uncomfortably existential questions are raised. While demand management is the immediate issue, the underlying challenge is safeguarding the power to choose.

This is not just a funder problem – or at least, other versions of the same problem are present as chokepoints across the whole research ecosystem. For academic researchers, the issue is information overload and the need to keep pace with the rise in the amount of published research – attributable to AI-assisted writing and factors like the exponential growth of newer research nations (China, India etc.), the global spread of open access publication models and the proliferation of “publish or perishopens in new tab/window” culture. For publishers, the challenge is assimilating all this information while retaining a meaningful gatekeeping function. For industry the hiring crisis, resulting from the huge mass of AI-supported applications, is making the recruitment process more time and resource intensive. Finally, for many governments, the bottleneck is in public regulatory and policy consultation, where AI has dramatically lowered the barrier to generating long, technically sophisticated submissions. This impacts the research ecosystem because consultations on R&D strategy and areas like research ethics and intellectual property directly determine which fields get resourced. In each case, a selection mechanism that depended on volume being naturally limited has been structurally broken by the rise of AI. In other words, the technology that many believe should be heightening the performance of the system is threatening to bring it to a standstill.

Adjudicators to orchestrators

Like researchers, publishing executives, hiring managers and civil servants, funders are rapidly working to modify their selection and evaluation processes. The first impulse has naturally been to implement curbs on the use of AI in grant proposals, but these are difficult to enforce, while even limited applications of the technology, like those permitted by UKRI and JST, may do little to abate the flow of submissions, with the general uplift in quality levels ultimately making decision-making much harder.

Given the developments around AI detection, triaging and streamlined peer review discussed above, it seems safe to assume that the future will involve more rather than less AI, with the crisis providing a catalyst for innovation, albeit with expert human decision-making remaining front and center. Indeed, a heightened sense of the limitations of AI may be the real key to success, with funders compelled to tackle longstanding problems associated with fragmented infrastructure and limited access to external data sources in order to realize the full benefits of the technology. That said, robust and responsible AI systems offer a real opportunity to remake the funding system, ensuring grant allocation processes are faster and fairer, broadening and accelerating discovery and supporting public accountability through more accurate impact reporting.

In the long term, AI may also prove to be a powerful connector across the research ecosystem, with funders, government, academia and industry historically struggling to maintain real-time visibility into one another’s priorities. Arguably, this is hampering their ability to respond effectively to the needs of taxpayers, especially around grand challenges like pandemic preparedness or net zero – or, ironically, the rollout of new technologies like AI itself.

Moreover, this new infrastructure could see funders shift towards a more proactive, intelligence-led operating model, anticipating where investment will have the greatest impact and then brokering the relevant connections between partners to make it happen. Empowered by AI, perhaps public funders will no longer be passive adjudicators of competing grant proposals but active orchestrators of the entire innovation ecosystem.