AI drives European grant surge, GPT-4o matches Austrian reviewers

Surge in AI-Generated Grant Submissions: Funding agencies across Europe and the Netherlands report higher application volumes attributed to text generation software. Because grant allocations remain static, acceptance rates have fallen, and review panels note instances where applicants cannot explain the experimental procedures detailed in their written texts during oral evaluations.
Language Model Agreement with Grant Panels: An evaluation of GPT-4o against historic Austrian Science Fund records showed that the model matched human reviewer funding determinations in a majority of cases, while divergence occurred across the remaining submissions.
Proposal Editing Tool Grounded in Citations: The Allen Institute for AI and the University of Pennsylvania created LitPivot, a tool that divides research drafts into functional sections to supply literature-based revisions during proposal formulation.
Data Governance in Human Subjects Operations: Researchers at Michigan Technological University assessed automated workflows for literature collection and survey selection, noting risks of data leaks and factual errors when handling participant records subject to Institutional Review Board oversight.
Automated Peer Review Support: Springer Nature and Google DeepMind are constructing a prototype system to provide workflow assistance for human reviewers evaluating academic manuscripts.
Additional Developments: A 2026 employment report reviewed the effects of automated writing software on grant administration personnel; OpenAI announced GPT-6.1 Terra; Google released Gemini 4 Carbon; and the UK Government published progress reports on its artificial intelligence programs.

Sources

Researchers face record-low chances of getting a grant partly due to rise of AI
Science grant providers are experiencing a surge in applications partly driven by the rise of artificial intelligence, which makes it easier to quickly create polished proposals. With the number of available grants remaining fixed, acceptance rates have dropped to record lows across major Dutch and European funding programs. Reviewers note that while AI-crafted submissions often appear well-written, some applicants struggle during follow-up interviews to explain how they actually intend to carry out their research.
AI Model Matches Human Grant Decisions 60% of Time | Wisevoter
A study found GPT-4o agreed with human reviewers on Austrian Science Fund funding decisions in most, but not all, cases.
AI2's LitPivot Helps Researchers Refine Ideas With Literature
The Allen Institute for AI and University of Pennsylvania researchers have developed LitPivot, a system that segments research proposals into facets and suggests literature-backed revisions as ideas evolve. A lab study found it outperformed standard AI chat tools on literature grounding.
Using AI responsibly in human research
Michigan Tech researchers are exploring how artificial intelligence can be utilized effectively in research while addressing critical confidentiality and safety concerns associated with human-derived data. While AI can aid researchers in selecting validated surveys and conducting literature reviews, using it with participant data presents risks of privacy breaches, unauthorized data exposure, and inaccurate model outputs. Consequently, researchers and students are urged to adhere to Institutional Review Board guidelines, protect participant information, and fully comprehend AI-generated…
Springer Nature and Google DeepMind to build 'peer review assistant'
Publisher and AI giant working on a prototype to explore how AI can support reviewers

Also this week

Full transcript

A surge in automated writing tools is driving grant proposal submissions across Europe to record levels. How funding agencies handle that influx is what we examine today on AI in RA, tracking artificial intelligence across research administration. Here is the reporting. Funding bodies across Europe and the Netherlands are seeing a direct consequence of automated writing tools: a steep rise in grant applications. And because the allocated funding pools have not grown, that surge in submissions pushes acceptance rates down to record lows across several major programs. What selection panels report seeing on paper is structured, clear text. The friction appears when applicants arrive for in-person interviews and struggle to explain their underlying experimental designs. The written draft conceals gaps that show up once someone has to defend the methodology out loud. At the same time, agencies are testing whether language models can sit on the other side of that process and mirror human funding choices. Researchers tested that scenario directly using GPT-4o on historical proposal outcomes from the Austrian Science Fund. The model matched the human panels on a majority of the grant determinations, but diverged on the remainder. It documents both the reach of automated scoring and its clear boundary lines. Other initiatives are targeting specific stages of proposal preparation rather than final scoring. The Allen Institute for AI and the University of Pennsylvania developed LitPivot for that purpose. LitPivot partitions a draft into functional facets, tracking conceptual shifts and recommending citations as the outline develops. In lab evaluations, it produced higher literature grounding scores than standard conversational interfaces. On the evaluation side, Springer Nature and Google DeepMind are building a prototype tool for peer review. The stated scope there is assisting referee workflows, rather than substituting automated output for editorial judgment in final decisions. Introducing external models into research workflows also creates data governance problems, particularly with human subjects. A review from Michigan Technological University examined tools applied to survey selection and literature synthesis. They tracked vulnerabilities tied to running participant records through third-party systems, documenting risks of data exposure, privacy breaches, and unverified errors. Their protocol guidance calls for following Institutional Review Board procedures, keeping raw subject data off external servers, and requiring human verification for every output. On the radar to close out this cycle: a 2026 workforce report evaluates the effect of artificial intelligence on grant writing and nonprofit positions. OpenAI has announced GPT-6.1 Terra, Google has announced Gemini 4 Carbon, and the UK Government released updates on its artificial intelligence initiatives. We will return next week with additional coverage on machine systems in scientific workflows. Until next time, on AI in RA.

From our community

Builders Group

A peer working group for research administrators who are actively building AI into their offices. Some members work alongside IT teams; others are piecing it together on their own. What unites the group is the work of implementation itself:

RSVP →

From our community

Join AI in Research Administration

Join the community →