UK Research and Innovation (UKRI), the UK’s umbrella funding agency, is publishing data underlying up to 2000 grant proposals to investigate whether the use of generative AI can reduce the burden of grant peer review.
UKRI allocates more than £8 billion to research funding each year. The number of research and innovation grants UKRI funds has halved over the past seven years, but the number of applications has soared by more than 80%.
So the agency is looking at ways to streamline the peer review process. In October, a research team led by Mike Thelwall, a data scientist at the University of Sheffield, began looking at how UKRI could use generative AI. The research was funded by the UK Metascience Agency, the first government agency dedicated to studying how research is done and how that process can be improved.
Mr Thelwall said his team would have access to full text versions of between 1,000 and 2,000 grant proposals submitted to UKRI that were ultimately either funded or rejected by UK research institutions, which are normally kept confidential. Thelwall and his colleagues plan to run these applications through a large-scale language model (LLM) to see if the tool can accurately predict the score reviewers give a proposal and the final recommendation decision.
Thelwall and his team will know the score each proposal received and whether it receives funding, but they will not disclose this to the LLM. “If a large-scale language model can do some reasonable job of predicting the score that a grant proposal will receive, then it may be possible to use language models in some way to speed up the grant review system or support the work of reviewers,” Thelwall says.
Mr Thelwall was previously part of a team that looked at how AI could be used to help assess papers submitted to the UK’s Research Excellence Framework, which assesses the quality of research conducted at UK universities.
When Thelwall and colleagues published data in December 2022, they recommended that more work is needed for AI systems to assist with peer review, and the data suggested that AI systems produced the same scores as human reviewers 72% of the time. However, Mr Thelwall said at the time that this figure would need to reach 95% accuracy.
Mohammad Hosseini, who studies the ethical implications of AI at Northwestern University in the US, said there were still “serious questions” about whether LLMs would generate novel ideas. “If AI cannot generate truly novel ideas, it is also unlikely to detect truly creative ideas because it is trained on existing data,” he says. “In a manuscript, you’re reporting what happened, but in a grant, you’re sharing ideas that might still be possible.”
Another problem for funders using LLMs, Hosseini points out, is that without transparency about what criteria they are inputting into the AI, there will be pushback from researchers. However, if funders are open about the process, grant applicants may begin to game the system or intentionally write in a way that may generate more favorable feedback from the AI.
It’s unclear how UKRI will use generative AI, but Terwall suggests it could work well in tie-breaking situations. Thelwall suggests that LLMs could also act as a third or fourth additional reviewer or assist with a quick desk rejection option to reduce the amount of peer review performed by human experts.
Terwall cited the example of La Caixa Foundation in Barcelona, which is experimenting with AI-powered peer review of grants. Thelwall said about 90% of submitted grant applications still undergo a full peer review by three experts.
“This saves all the reviewers a little bit of time, which doesn’t seem like a big deal, but it means that a lot of experts don’t have to spend time evaluating proposals that have a very low chance of getting funded,” Thelwall said.
