AI tool paints bigger picture of global impact of millions of research grants, papers and patents

Every year, billions are ploughed into research funding around the world that fuels scientific discovery. But where does that money actually go? Scientists have created a new machine learning AI tool called Funding the Frontier (FtF) that maps the impact of research grants onto publications, patents, policies, clinical trials and even news stories – providing a glimpse into how funding really shapes science and society.

Although research has long stressed the importance of science funding, it largely centres on grants and papers, leaving other kinds of impact underexplored. FtF is designed to help funders, policymakers, university leaders and researchers see the bigger picture by showcasing how science moves from investment to innovation in an informed and transparent way and is outlined in a preprint that has not yet been peer reviewed.

Built on one of the largest datasets of its kind using global data collected from platforms like Dimensions, SciSciNet, and Altmetric, FtF connects over 7 million research grants to 140 million scientific papers, 160 million patents, 10.9 million policy documents, 800,000 clinical trials and 5.8 million news articles – all linked through 1.8 billion citation relationships.

‘FtF was designed through close collaboration with real-world decision-makers,’ says first author Yifang Wang at Florida State University. ‘[It] could enable a shift … to a comprehensive, multidimensional view of impact. Decision-makers could see not only which projects produce papers, but also which ones contribute to innovation, policy, health or public understanding.’

‘FtF also visualises funding distribution at multiple levels: by field, institution, gender and career stage, which allows users to see who receives funding and where potential inequities exist,’ she adds.

This could revolutionise funding decisions by prioritising high-impact research and better forecasting future opportunities. ‘[It is] an ambitious and laudable attempt to combine and synthesise a wide range of data, indicators and algorithms,’ says James Wilsdon at University College London, who was not involved in the study. ‘What makes this different from most scientometric papers is the number of different elements being combined in a single … framework.’

‘The main strength of the system is a promising user interface of the integrated data sources that allows users to explore scientific outputs, their outcomes and impacts, in what seems to be an intuitive way,’ adds Vincent Traag from Leiden University, also not involved in the study.

Worries that Funding the Frontier might push funders towards ‘safe’ research

But critics warn that overreliance on metrics and predictive models could bias funding toward ‘safe’ projects and undervalue long-term or curiosity-driven science.

‘Could there be a risk of research becoming a self-fulfilling prophecy if we make decisions based on such algorithms?’ asks Wilsdon. ‘Could decision making become more conservative, if based on what worked in the past, not what is expected to work in the future?’

That worry is shared by ethicists, who caution that using past performance as a guide for future funding could entrench the status quo. ‘When a system like this is used [to allocate] funding to research that has, in the past, yielded [impact], the potential harm is that … it might lead one to avoid funding innovative research that generates new types of impacts [or] break historical patterns,’ comments Philip Brey at the University of Twente in the Netherlands, who was not involved in the study. ‘A lot of innovative, groundbreaking and disruptive research does not immediately generate impact, but is nevertheless worth funding.’

Experts in AI policy also warn against overconfidence in FtF’s results – trusting the outputs without considering how they were produced or the uncertainty behind them. Traag argues that we tend to scrutinise human-made judgments more readily, however, when it comes to algorithmic predictions, it’s easy to just accept them. ‘Yet there is still considerable uncertainty in predicting which grants will generate which impacts … and people may place too much confidence in the results,’ Traag adds.

However, FtF’s predictions are based solely on patterns in grant text, making its predictions inherently partial. ‘This is a rather straightforward translation of words into impact without further and broader contextual information,’ says Traag. ‘A more useful perspective would probably be to sketch a grant proposal abstract in a broader context to allow users to explore … where the proposal goes beyond existing knowledge.’

The consensus is that, in practical decision-making, systems like FtF should serve only as a supporting tool, rather than directing outcomes. ‘These types of predictive analytics need to be approached with caution, and tested thoroughly, before being adopted and implemented by research funders and policymakers,’ says Wilsdon.