In the latest episode of the NCN podcast, Prof. Margaret Ohia-Nowak and dr inż. Tomasz Szumełda discuss the results of a survey on the use of AI in proposal writing, as well as research into bias in Polish language models—a project led by Margaret Ohia-Nowak and funded by the NCN. Hosted by Anna Korzekwa-Józefowicz.
Margaret Ohia-Nowak is a linguist and media scholar at the Maria Curie-Skłodowska University in Lublin. Her research focuses, among other things, on race, racialisation and the language of public discourse in Poland and Central and Eastern Europe. Tomasz Szumełda served for many years as an NCN scientific coordinator. He is currently involved in implementing the e-Grants system (e-Granty) and represents NCN in the AI working group at Science Europe, an association of research funding agencies and scientific institutions.
AI in the grant system
In autumn 2025, NCN conducted a survey among principal investigators of projects submitted in recent editions of the OPUS, SONATA and PRELUDIUM calls, asking about the scope of GenAI use in preparing proposals and the limits of its acceptable use. A total of 2,708 respondents took part. Around 60% of respondents believe that NCN should allow the use of GenAI in preparing proposals, while 42% admit they have already used such tools. At the same time, most respondents oppose the use of AI to develop research concepts, formulate hypotheses and prepare literature reviews. The position is even clearer when it comes to proposal evaluation: 67% of all applicants and 72% of grantees exclude the use of GenAI in this process.
Tomasz Szumełda notes that the survey results are consistent with NCN’s position published in May 2025, which allows the auxiliary use of GenAI in preparing proposals and prohibits its use in scientific evaluation. “The research community itself separates form from content, which shows that NCN’s approach aligns with applicants’ expectations,” he says. Margaret Ohia-Nowak, who also completed the survey, agrees with this distinction and adds that AI can be useful for translation and language editing, but the research concept must remain the work of the researcher.
Tomasz Szumełda also emphasises that the guidelines developed by NCN are aligned with a broader European approach. Work is ongoing within Science Europe to develop common recommendations for national agencies, and NCN’s current approach—allowing the auxiliary use of AI by applicants and prohibiting its use in evaluation—is consistent with the prevailing direction in Europe. An exception is the German agency DFG, which has recently allowed reviewers to use AI for language editing and structuring their own critical comments. However, DFG prohibits the use of publicly available models, allows only local institutional servers, and requires reviewers to indicate which parts of their reviews were generated using AI.
Subtle algorithmic bias
The podcast also discusses AI as a field of research. NCN funds projects on AI across multiple areas, from machine learning algorithms and language models, through applications in medicine and biology, to analyses of its social and legal implications. One such project is Margaret Ohia-Nowak’s research on Polish large language models (LLMs). The researcher analyses how the teams developing the Bielik and PLLuM models address safeguards against reproducing ethnic and racial stereotypes. Findings from the first phase indicate that Polish models include a wide range of bias-mitigation measures, and that the range of solutions continues to expand.
The researcher notes, however, that initial, instinctive model responses may be less nuanced, as illustrated by the following example:
“But even today, I asked ChatGPT to generate an image of a wise person. And that wise person… White, with a moustache. These models are now able to differentiate. But three years ago, in 2023, when the idea for this project first emerged for me, they were not. From my corpus research, I already know that these models can be subtly biased. Even when trained to reduce bias, their first response to a question about a ‘wise person’ often reflects a very stereotypical image. The question is: how can we make those initial responses different?” she says.
The project, scheduled for completion at the end of 2027, is expected to result in a guide for two groups: users, supporting them in constructing prompts that generate more culturally sensitive content, and teams developing future Polish LLMs, providing them with a set of good practices.
Selected statements
Margaret Ohia-Nowak
I smiled when I heard that most respondents do not support using AI tools to develop research concepts. I also selected that answer. However, when it comes to formalities, which are often difficult—especially for those with no experience in writing proposals, or in structuring them—this can also be helpful, although much of it is already specified in the call or proposal description. I think these are technical aspects that funding institutions should allow, and NCN does. It involves support with minor translations or language editing, but not with the creative contribution.
Tomasz Szumełda
At this point, NCN’s position is clear: AI should not, in any way, interfere with the merit-based evaluation process. We could probably record a whole separate podcast on this topic. Let’s imagine, for example, an AI reviewer who, having been fed a huge database showing that this type of research has received funding over the last three years, might conclude that the research is not innovative—or, conversely, that it is innovative—because it will have been fed a dataset that subsequently suggests appropriate responses.
There have been situations where experts and reviewers evaluating the proposals did indeed feel that something was off. That the research concept or the proposal was in some way superficial, generic. Someone asked a simple question and received a simple answer. What was missing was depth and coherence between the methodology and the track record. Such proposals were typically evaluated as being of low quality, yet they required the involvement of human resources to assess something that had no substantive value.