Towards responsible use of AI in research
The rapid development of artificial intelligence not only contributes to improving efficiency in research and its funding but also opens new pathways to scientific insights. However, the risks associated with AI must also be considered.
Researchers and funding agencies are increasingly using tools based on artificial intelligence (AI). AI tools help researchers to process large volumes of data, identify new correlations and carry out tasks such as research, proofreading and translation more efficiently. The transformative nature of AI can impact numerous scientific disciplines.
The SNSF is aware of both the potential of AI systems and the unresolved questions and risks associated with their use in science. The latter include, for example, maintaining good scientific integrity and ensuring the traceability and reproducibility of results achieved with AI. Addressing these questions is itself a subject of research funded by the SNSF.
AI as an umbrella term
Artificial intelligence encompasses a wide range of technologies and methods. These include approaches that have been researched intensively for decades, such as machine learning (statistical methods that recognise patterns in datasets to make predictions based on them), neural networks (approaches that mimic the functioning of the human brain) and natural language processing (the processing and understanding of human language). Simulating human intelligence is just one of many possible goals and applications.
Public debate currently centres on so-called generative AI technologies. These technologies create content such as text, images or videos by using text commands (prompts). Researchers also employ these methods across various disciplines for many different tasks, such as analysing and processing data and identifying patterns and correlations in datasets.
The responsibility remains with researchers and reviewers
The SNSF welcomes researchers harnessing the potential of AI for their work. Anyone who uses AI in research remains responsible for the results and must adhere to the principles of scientific integrity and the protection of confidential data.
The same principle also applies to funding proposals and their evaluation: Researchers, reviewers and referees remain responsible for all content, assessments and decisions that they produce with the support of AI.
The SNSF has updated its guidelines for reviewers and referees. These guidelines set out the purposes for which reviewers and referees may use AI and the conditions that apply.
Reviewers must first form their own independent opinion. They assess the quality of a proposal and the applicants’ competences. Afterwards, AI tools can, for example, help to improve language or assist with literature research. Expert reviewers must maintain the confidentiality of the data entrusted to them. They may only use AI tools whose settings and options ensure that any uploaded content is neither permanently stored nor used for training purposes. When reviewers submit their assessments, they must disclose to the SNSF whether and how they have used AI.
The SNSF already uses two key AI technologies to process funding proposals: natural language processing and machine learning techniques. The SNSF Data Team has developed an approach to assist employees in assigning funding proposals to reviewers with the necessary expertise. The application, which we are currently testing in day-to-day operations, analyses textual similarities between excerpts from the proposals and the publications of the experts. Based on this analysis, the system suggests potential reviewers. The suggestions are then reviewed and, if necessary, adjusted by SNSF staff.
In addition, the SNSF is collaborating with international research funding organisations to explore how the processing of funding proposals could be improved with the help of AI applications. This includes participation in the project GRAIL ("Getting Responsible about AI and machine learning in research funding and evaluation") by the Research on Research Institute (RoRI). The project, which began in mid-2023, concluded in 2025 with the publication of a handbook on responsible uses of AI in research funding.
Links
- Research on Research Institute
- GRAIL project
- Handbook for responsible uses of AI and machine learning by research funders
Year
FAQ
Reviewers and referees remain fully responsible for their assessments, which must be based on their own expertise and a careful reading of the proposal, and, in the case of referees, the external reviews. AI tools may only be used in a supportive role after reviewers and referees have read the proposal and formed their own (preliminary) opinion. In practice, the use of AI is allowed for refining and expanding on one’s own assessment of the proposal. It is not allowed to use AI to generate parts of or the full assessment of the proposal.
Substantial use of AI means any use that generates core elements of the review – such as the criteria-based assessments, strengths and weaknesses, or the overall assessment of the proposal – without reflection by the reviewer or referee. Generating substantial parts of a review without independent assessment by the reviewer or referee is prohibited. For example, prompting an AI system to generate a full review of a proposal or parts of it is considered substantial use and is not permitted.
Acceptable uses are limited to supporting tasks, such as improving language, translating, structuring one’s own and/or reviewer-written notes, assisting with literature searches, or refining, verifying and/or expanding upon one’s own assessment notes. The intellectual content of the assessment must always originate with reviewers and referees.
Yes. Any use of AI tools in preparing the assessment, even in a compliant supporting role, must be disclosed when submitting the review.
Only if the AI tool guarantees full confidentiality, data protection and no access by third parties, in accordance with the SNSF’s evaluation guidelines. (PDF) The AI tool may not use the contents of the proposal materials for model training or store data indefinitely. Most publicly available AI tools do not meet these requirements by default and need to be manually adjusted.
Institutionally hosted AI tools may be used if they meet strict confidentiality and data protection requirements. The AI tool, regardless of whether it is internal or institutionally hosted, may not use the contents of proposal materials for model training or store data indefinitely. The key criterion is not the type of tool, but whether sensitive data remains fully protected and only accessible to the reviewer or referee.
Yes, if they comply with the SNSF’s evaluation guidelines, locally hosted AI tools without internet access may be used in a supportive manner for tasks such as improving language, translating, helping with literature review, refining and expanding on one’s own assessment notes, etc. Reviewers and referees remain fully accountable for the output.
The main risks are a loss of intellectual ownership of the assessment and potential breaches of confidentiality. Both would undermine the integrity of the evaluation process.
No. Even if the output is reviewed afterwards, generating substantive parts of the assessment with AI without prior assessment by a reviewer or referee (e.g. in the form of bullet points) is not allowed. The evaluation must remain under the full intellectual control of reviewers and referees. For example, using AI is allowed for refining and expanding upon one’s own assessment of the proposal. It is not allowed to use AI to fully create an assessment of the proposal or parts of it.
Yes, this is allowed, as long as the AI tool guarantees full confidentiality, data protection and no access by third parties, in accordance with the SNSF’s evaluation guidelines (PDF).
This is allowed, as long as the AI tool guarantees full confidentiality, data protection and no access by third parties, in accordance with the SNSF’s evaluation guidelines (PDF).
No, this is not allowed. The assessment must be based on the information provided by the applicants and the SNSF-supplied documents. The information requested from the applicants and included in the official documents for review constitutes an integral part of the evaluation procedure.
Yes, this is allowed, as long as AI tools are used to support the reviewers or referees’ own expertise rather than to replace their assessment. For example, a legitimate use of generative AI could be to request that the AI tool check for existing published research related to the uploaded proposal’s questions.