Appendix A – Guidelines, risk management

Artificial intelligence:

• Has no sense of morality
• May perpetuate or amplify biases found in training data (e.g., gender, ethnicity, etc.)
• May give false or misleading information ("hallucinations")
• May base results on outdated, incomplete, or unreviewed data that can provide unreliable information about the latest developments or facts

Users shall therefore:

• Criticize and verify: Verify all information, statistical findings, and references with trusted, up-to-date sources
• Ensure quality and coherence: Ensure that outputs are logical, coherent, and meet the professional and competency criteria of the project
• Correct errors and biases: Be aware of errors, systemic biases or possibly outdated knowledge and correct them
• Take responsibility: Take full responsibility for the content and ethical aspects of the material, as if it were your own work

Such contributions include:

Initial conceptual work and definition of a research question or objective.

Development of one's own reasoning, interpretation, or analysis.

Organization and structure of the project.

Design of targeted and thoughtful prompts.

Critical selection, processing, and integration of AI output.

Interpretation of data or results, even if AI has assisted in data processing.

Not considered a significant contribution:

• Using only generic or unelaborated prompts for AI without supporting them with independent analysis, processing, or thoughtful changes to the model's results.
• Copying and returning raw, minimally altered or unreviewed AI outputs and presenting it as their own work.

A more detailed description is necessary if:

• AI had a significant impact on results or methodologies.
• It was used to write code, perform data processing, create visuals, or develop core arguments.
• Results are based on AI outputs to the extent that reproducibility may require prompts, implementations or settings to be recorded and delivered as an appendix, in a methodology chapter or in supplementary data, as appropriate. This is especially true in research and projects where transparency, methodology and the possibility of repetition are important.

The goal is to ensure that evaluators can understand, evaluate and, if necessary, replicate the processing process and results.

Note: The level of transparency shall be comparable to that required for the use of specialised tools (e.g. statistical programs) or the assistance of colleagues. Special requirements of courses and project guides take precedence.

Purpose

UNAK shall carry out regular risk assessments for new or changed artificial intelligence tools used in education, teaching, research, support services or administration. The goal is to promote a responsible implementation process in line with the policy's emphases and international norms, especially requirements according to the EU AI Act (2024) and GDPR.

Description of the risk assessment

UNAK shall use a simple and efficient risk assessment model when new tools are introduced or significant changes are made to existing systems.

Six main categories shall be examined and evaluated on a scale from 1 (low risk) to 5 (very high risk):

When assessing risk, the following shall also be considered:

• The nature and extent of use (e.g. in teaching, administration, research).
• Potential risk mitigation (e.g. through human supervision, clear procedures, restriction of access, etc.).

The risk assessment shall be documented and maintained by the AI Project Manager.

Decision and measures

If a risk assessment reveals a high or very high risk (level 4 or 5 in any category), the following shall:

• Carry out a more detailed assessment in consultation with UNAK's security and privacy team.
• Ensure clear management approval before implementation.
• Establish clear rules for use and human supervision.

If the risk is considered low to medium (levels 1–3), implementation can proceed according to the general processes and conditions of the policy.

Treatment of suspected abuse

If there is suspicion that a student has used artificial intelligence in an unauthorized manner, the rules that apply to violations of academic integrity at the University of Akureyri shall be followed.

The University of Akureyri uses automatic analysis tools (e.g. Turnitin) to detect possible plagiarism. However, the additional functionality of such systems to detect the origin of text with respect to AI is considered unreliable and should therefore not be used as a basis for decisions on violations related to the use of AI.

Research (Mobin & Islam, 2025) supports this precautionary view and shows that even the most powerful systems can be unreliable, especially when texts are from different sources or models. Therefore, the results can be misleading (false positives) or omit actual fractions (false negatives).

Emphasis on professional evaluation

Instead of relying on automatic technology, the situation should be assessed with professional insight and direct dialogue with the student, e.g. through an oral review or questions about the work process. Trust, professionalism and consistency shall be the guiding light in the handling of the case.