CUNI: Recommendations for Research and Researchers (full text)

Source: https://ai.cuni.cz/AIEN-14-version1-ai_3_en.pdf Issued by: Inter-university working group (Charles University, Masaryk University, Czech Technical University, and others) Status: Freely distributable; edited and updated regularly Note: MUNI co-authored this document and endorses its principles, even though MUNI has not published a standalone researcher AI policy of its own.

Opening statement

"The democratization of AI technologies does not change the basic principles of research work, not only in the area of the research itself, but also with respect to preparation and administration. Research should continue to be transparent in its methods and ethical in its execution and should strive to maximize reproducibility. When conducting research, researchers should act in accordance with the FAIR principles as much as possible."

Four principles

1. Transparency

"Artificial intelligence is a tool for conducting research, not the author of the research."

• AI-generated results may yield new discoveries or unexpected connections, but AI's role is primarily heuristic — it cannot author research interpretation

• AI participates in research similarly to applied statistics or data visualization

• The results and interpretation of research are always the responsibility of the researcher

• Researchers should specify in their research methodology "in a transparent manner how and where AI has been used with a view to future research replicability"

• Ideal practice: each research result should be accompanied by "a separate file with outputs, prompts, and model settings"

2. Ethics

Transparency in AI use relates to ethics on two main points:

1. Origin of training data — training data should be protected against exposure of sensitive data in the model
2. Distortions in data — "The behaviour of the models should then be tested for the presence of biases — gender, racial or other, which could be implicitly hidden in the nature of the data used"

The researcher has full responsibility in the area of ethics.

3. Reproducibility

• Publish datasets and source codes used for training and testing AI, "as long as it is not in conflict with the protection of personal data and intellectual property"

• If your own or a trained model is used: publish the model together with the code, data, and exact parameters used for training

• For qualitative research specifically: publish the prompts or codes used for coding qualitative data

• For generally available models: provide "complete and accurate information about parameter settings, such as hyperparameters, which may influence research results"

4. AI's Assistant Role

AI has a key role in assistance tasks beyond direct research:

• Can automate text generation for project design (grant proposals, etc.), "allowing researchers to focus more on the research itself than on the excessive administration surrounding it"

• Can automate routine data analysis tasks: "cleaning, transforming, and visualizing datasets"

Warning about inequality:

"Intensive use of AI tools can create benefits for those who use them and widen the gap between researchers and workplaces that use AI and those that do not."

The document calls on research institutions to "support their researchers in acquiring the skills to work with AI and to provide them with adequate facilities and resources."

What this document is and is not

Is: The most detailed researcher-facing AI guidance currently available from Czech universities. Freely distributable. Co-authored by MUNI.

Is not: Binding regulation. Does not address: specific journals' requirements, grant funder rules (GAČR, ERC), authorship in publications, or peer review.

Practical implication for DISSINET researchers: Following these four principles covers the institutional dimension. For publication and grant requirements, follow the relevant journal/funder policy (see E.journal-funder-policies when available).

Related

• E.muni-guidelines — MUNI's own (non-binding) student and teacher guidance

• A.critical.limitations — the Pepinsky execution/interpretation distinction maps directly onto "AI is not the author"

• A13.examples-dissinet-usecases — data sensitivity and qualitative research cautions