Imagine a world where groundbreaking research ideas are drowning in a sea of paperwork, with funding decisions taking forever to make— and that's exactly the crisis facing the UK's top research funder right now. But here's where it gets controversial: what if artificial intelligence could swoop in to save the day, or is it just a risky gamble that might stifle true innovation? Let's dive in and explore this thrilling yet divisive topic together.
The UK's primary research funding organization, UK Research and Innovation (UKRI), is pioneering a bold experiment by releasing sensitive data from up to 2000 grant proposals. Their goal? To investigate if cutting-edge generative AI can lighten the load on the time-consuming process of grant peer review—a system where experts evaluate and score research proposals to decide which ones get funded. For beginners, think of peer review as a quality check: it's like having a panel of knowledgeable judges review your business pitch to see if it deserves investment, ensuring only the strongest ideas get the financial backing they need.
UKRI doles out over £8 billion annually to fuel scientific and innovative projects. Yet, while the actual number of grants they've approved has dropped by half in the past seven years (as detailed in this Nature article: https://www.nature.com/articles/d41586-025-02584-w), the flood of applications has skyrocketed by more than 80%. This surge creates a bottleneck, overwhelming reviewers and slowing down progress. So, in a bid to modernize, UKRI is exploring smarter ways to streamline their review process.
Come October, a team headed by Mike Thelwall, a data science expert at the University of Sheffield in the UK, will kick off this initiative (more on their project here: https://sites.google.com/sheffield.ac.uk/llmgrants). Their work is backed by the UK's first government-funded Metascience Unit, which focuses on studying and enhancing how research itself operates. Thelwall's group will get their hands on the complete text of 1000 to 2000 grant proposals that were either successfully funded or turned down by UKRI—information typically kept under wraps for confidentiality.
The plan? They'll feed these proposals into powerful large language models (LLMs), which are advanced AI systems trained on vast amounts of text to generate predictions. For those new to this, LLMs are like super-smart chatbots that can analyze patterns in language to make educated guesses, similar to how a seasoned editor might predict which manuscript will become a bestseller. The team aims to test if these AI tools can reliably forecast the scores assigned by human reviewers and the ultimate funding recommendations.
Importantly, Thelwall and his colleagues will know the real scores and outcomes for each proposal, but they won't feed that info into the LLMs. As Thelwall explains, 'If large language models can do some kind of reasonable job at predicting the score that a grant proposal would get, then that might allow them to be used in some way to help speed up the grant review system or to support the work of reviewers.' And this is the part most people miss: it's not about replacing humans entirely, but augmenting them to handle the workload more efficiently.
This isn't Thelwall's first AI rodeo in the world of academic evaluation. He previously contributed to a study examining how AI could aid in reviewing articles for the UK's Research Excellence Framework, which ranks the quality of university research across the country (check out the Nature piece here: https://www.nature.com/articles/d41586-022-04493-8). Back in December 2022, when they shared their findings, they concluded that AI still needed improvement before it could reliably assist in peer reviews. Their data showed that AI matched human reviewers' scores only about 72% of the time, far short of the 95% accuracy threshold Thelwall deemed necessary for practical use. It's a reminder that while AI is powerful, it's not infallible—much like relying on a GPS that sometimes takes you on a scenic detour.
But here's where it gets controversial: critics like Mohammad Hosseini, an expert on AI ethics at Northwestern University in the US (find his profile here: https://www.feinberg.northwestern.edu/faculty-profiles/az/profile.html?xid=62124), raise serious concerns. He argues that LLMs struggle to generate truly original ideas because they're trained on existing data. 'If AI cannot create really novel ideas, it also is unlikely to detect really creative ideas because it is being trained on existing data,' Hosseini points out. To clarify for newcomers: in a published paper, you're describing experiments already conducted, but a grant proposal is about pitching future possibilities—bold visions that haven't been tested yet. AI, drawing from the past, might overlook groundbreaking concepts that push boundaries.
Transparency is another hot-button issue. Hosseini warns that if funding bodies aren't upfront about the criteria programmed into the AI, researchers could rebel against what feels like a black box decision. On the flip side, full openness might lead to 'gaming the system,' where applicants tweak their proposals to align with AI preferences, potentially prioritizing buzzwords over substance. Is this the start of an arms race in grant writing, or a fair evolution in evaluation? And this is the part most people miss: could AI inadvertently favor certain styles of writing, sidelining diverse voices or innovative but unconventional ideas?
While UKRI hasn't spelled out exactly how they'll implement generative AI, Thelwall envisions practical applications that could revolutionize the process. For instance, AI might excel in close-call scenarios, acting as a tiebreaker when human reviewers are split. It could also serve as an extra set of eyes, like a fourth or fifth reviewer, providing additional insights without overburdening experts. Alternatively, it might enable a quick desk-reject feature, weeding out low-potential proposals early to free up reviewers for the most promising ones. As an example, consider the la Caixa Foundation in Barcelona, which has piloted AI-assisted grant reviews in biomedical research (detailed in this Cambridge study: https://www.cambridge.org/core/journals/data-and-policy/article/aiassisted-prescreening-of-biomedical-research-proposals-ethical-considerations-and-the-pilot-case-of-la-caixa-foundation/D17CC86C884D4F9E3A6DD9ABF3F4605E). In their trial, about 90% of applications still go through full expert review, but AI helps save valuable reviewer time by triaging the rest. 'They save a little bit of full reviewer time,' Thelwall notes, 'which doesn’t sound like a lot but that’s a lot of experts not having to spend their time evaluating proposals which have a very low chance of being funded.' Imagine the hours reclaimed for actual research instead of paperwork!
As we wrap this up, ponder these questions: Do you think AI in grant review is a brilliant shortcut or a threat to creative freedom? Will it level the playing field or just amplify biases from the data it's trained on? Could it even discourage risk-taking in science by favoring predictable ideas? Share your thoughts in the comments—do you agree with Hosseini's skepticism, or are you optimistic about AI's potential? Let's continue the conversation; after all, the future of funding innovation depends on voices like yours.
