Prorector: Using AI to evaluate research funding applications could affect entire system

The Estonian Research Council is testing the use of artificial intelligence to evaluate research funding applications, prompting concerns from the University of Tartu over transparency, fairness and data protection.
The Estonian Research Council (ETAG) plans to test the use of artificial intelligence in evaluating research funding applications and has informed the University of Tartu of its plans.
In response, University of Tartu Vice Rector for Research, Professor Mari Moora told the council that while testing new technological solutions is a positive step, the experiment could affect the equal treatment of researchers, the transparency of the evaluation process and confidence in the research funding system. Its potential broader implications must therefore be considered.
"The letter sent to the research community does not make clear the nature of the planned activity — whether it is a research study, a technological validation test, an effort to develop evaluation methodology or an improvement to administrative procedures — or what specific task AI would perform in the proposed experiment," Moora said.
According to Moora, the letter also fails to explain how the use of AI would relate to the existing expert review process or what criteria would be used to evaluate the experiment's results. She believes several questions need to be clarified before the experiment begins, including how data protection and confidentiality will be ensured.
"At present, there is no information on whether the data will be processed on platforms operated by external AI service providers or in an environment controlled by ETAG, what contractual and technical safeguards will be used or how potential data protection and confidentiality risks have been assessed," the vice rector said.
She emphasized that research proposals may contain sensitive or unpublished information, making it essential to address these questions before the experiment begins.
Another concern is ensuring the quality of evaluations and equal treatment. Moora pointed out that the council has yet to explain how it intends to assess the overall impact of AI use on evaluation quality or how it will evaluate the risks of potential bias or unequal treatment.
"I believe it is particularly important to analyze the impact on different fields of research, research traditions and methodological approaches and to ensure that innovative, high-risk or unconventional research ideas are evaluated fairly even when AI is used in the evaluation process," Moora said.
Who will be responsible for AI evaluation results?
Accountability also raises questions. Moora noted that while human experts are personally responsible for their assessments, must declare that they have no conflicts of interest and risk their reputations if they act improperly, it is difficult to understand what considerations underpin an AI-generated assessment.
"If AI output influences a funding decision, it may become more difficult to assess whether that decision is justified and lawful. This, in turn, raises the question of how such a decision could be meaningfully challenged or reviewed in administrative proceedings, including how to establish whether errors of judgment or other procedural shortcomings occurred in the decision-making process," Moora said.
According to Moora, since the experiment is presumably intended to assess the potential use of AI in preparing research funding decisions, it should examine not only how the existing technological solution performs but also how it could affect the research system as a whole.
"Evaluation systems inevitably influence researchers' behavior, so attention should also be paid to the risk that researchers may begin tailoring their applications to the assumptions or expectations of algorithmic systems rather than presenting their scientific ideas in a way that best supports substantive evaluation by experts," Moora said, adding that this could limit research institutions' freedom to choose which areas of research to pursue.
The University of Tartu vice rector emphasized that her questions do not stem from opposition to the use of AI or efforts to improve research funding processes. However, she said these possibilities must be thoroughly examined, carefully considered and discussed with the research community.
The sustainability of the research funding system is another concern. Moora explained that generative AI could significantly increase the number of applications, while high-quality expert review depends on a limited pool of specialists in each field. Researchers therefore also consider it important to discuss how to maintain the quality of the funding and evaluation system if the volume of applications grows faster than the capacity to review them.
ETAg: Precisely the questions the experiment aims to answer
Estonian Research Council Director Kristiina Tõnnisson told ERR that the current phase could be described as a trial run for an experiment, using projects that have already undergone evaluation.
The council is now comparing AI-generated assessments of research projects with existing evaluations by human experts to establish a baseline for comparison.
Addressing the concerns raised by Moora, Tõnnisson said the experiment only involves projects whose principal investigators had checked with their teams to see whether anyone objected. No objections were raised. The evaluation experiment is being conducted on a couple of selected platforms. Although the council also considered developing an in-house system, it concluded that solutions developed specifically for Estonia performed considerably worse than large language models.
"So, for efficiency reasons, we decided to use large language models, but these operate in closed environments," she explained.
Regarding the risk that AI might favor familiar ideas and discriminate against original ones, Tõnnisson said the experiment is intended to examine precisely these kinds of risks. The trial should help identify which evaluation criteria AI struggles with, when it produces good summaries and when it does not.
"We hope to get more answers to these questions before we start saying that we are considering using AI to assist with evaluations in real-world situations," the council's director said.
The Estonian Research Council hopes to draw its first conclusions from the current experiment in about a month, before Christmas. It will then be able to consider whether this approach could be pursued further in the future.
--
Editor: Marcus Turovski











