AI Matches Grant Reviewers 60% of the Time
Early tests show GPT-4o agrees with human funders often enough to study, but it undervalues novel research.
Updated on Nov. 3, 2026 in Artificial Intelligence
Live Poll
Would you trust AI to help decide which scientific research gets funding?
A study published in Scientometrics in August found that GPT-4o agreed with human reviewers' funding decisions on 139 Austrian Science Fund grant proposals in 60% of cases. The research is at the study stage and does not describe a deployed grant-assessment system.
Why it matters
Funders are testing whether large language models can help triage research proposals, a task that is slow and costly for human panels. The 60% agreement rate suggests partial overlap with human judgment, while the model's reluctance to fund highly novel proposals points to a structural bias that matters for funding policy.
The study compared proposals in pairs within the same research field, using the prompt 'Which proposal should be funded? Answer with A or B only.' Proposals with the most pairwise wins were ranked highest, and the 60% figure reflects agreement with reviewers' final decisions.
The players
OpenAI
OpenAI, based in San Francisco, California, builds the GPT-4o large language model used in the study.
UK Research and Innovation
UK Research and Innovation is the UK public body that funds research and is exploring AI-assisted grant assessment through a controlled trial.
Wellcome
Wellcome, headquartered in London, is a major biomedical research funder that prohibits reviewers from entering grant applications into generative AI tools.
The details
GPT-4o is a large language model, a system trained on vast text to predict and generate language. Rather than scoring each proposal on its own, the study had the model compare proposals two at a time within the same field, and proposals that won the most head-to-head comparisons were ranked highest for funding. Each proposal carried human reviewers' assessments of novelty, scientific quality, team capacity and feasibility, which served as the benchmark. The researchers found the model tended to favor established ideas, because it learns from patterns that recur across its training data and therefore recognizes mainstream work more readily than genuinely new contributions.
Timeline
The 139 proposals were submitted to the Austrian Science Fund in 2021.
The study was published in Scientometrics in August 2026.
The Tech Race
UK Research and Innovation's planned controlled trial will compare AI-generated evaluations against past human funding decisions, a more rigorous test than retrospective agreement alone. The Austrian study gives that trial a baseline, showing where a general-purpose model currently diverges from reviewers.
Researchers applying for grants may see AI-assisted triage in future funding rounds if trials like UK Research and Innovation's proceed. Applicants whose work is highly novel should note the study's finding that the model was less willing to fund such proposals than human reviewers were.
The takeaway
Watch whether UK Research and Innovation's controlled trial reports agreement rates that exceed the 60% baseline in the Austrian study. The result will show whether AI-assisted grant assessment can move from retrospective comparison into live funding decisions.
Further reading
Explore more on Artificial Intelligence as models move into research evaluation and funding decisions.
Source note: This article includes information reported by Nature.
Live Poll
Would you trust AI to help decide which scientific research gets funding?
