Technical documentation

The AI tool uses information from several sources that are already used in the examination process – examination questions, examiner guidelines, student answers and grades – to create a first draft of the explanation.

The process takes place in several steps:

1. Access and authentication

Users log in via FEIDE to access the system securely.

2. Internal data processing

The documents are processed in a closed environment located in UiA's internal zone.

Here, relevant information is extracted from the documents (.pdf, .docx, .txt) before the data is prepared for use in the language model.

3. Draft generation

The system sends a structured request (prompt) to a language model (LLM – Large Language Model).

In spring 2026, Microsoft Azure OpenAI (ChatGPT 5) is used.

The language model then returns a draft explanation, which is made available to the sensor for quality assurance and editing.

4. Response and completion

The draft is displayed in a closed user interface.

The examiner reviews the text, makes any necessary academic adjustments and then sends the final explanation to the student via Inspera.

All data processing takes place within UiA's secure internal network.

Only persons who have access to examination data (e.g. examiners, subject teachers and the examination office) also have access when the AI tool is used.

• Data is not stored in the AI model and is not used for training.
• All processing takes place via internal APIs and through proxy servers that control the data flow between internal and external zones.
• The documents never leave UiA's secure environment.

A risk and vulnerability analysis (ROS) and a data protection impact assessment (DPIA) have been carried out, and the project has been discussed with the Norwegian Data Protection Authority. It has been agreed that the project has an acceptable risk profile and that the necessary measures have been implemented to prevent unintended incidents.

To ensure accurate and academically correct drafts, a standardised prompt is used that contains:

• instructions to the model regarding language, form and tone,
• placeholders for the assignment, examiner guidance, answers, grades
• and requirements for a clear academic explanation.

In addition, the language model is given access to the resources mentioned in How the solution works above.

The system prompt in v4.0 – spring 2026 – is as follows:

All students are graded according to this scale:

A Excellent - An outstanding performance that clearly stands out. The candidate demonstrates very high independence.

B Very Good - A very good performance. The candidate demonstrates very good judgment and independence.

C Good - Consistently good performance that is satisfactory in most areas. The candidate demonstrates good judgment and independence in the key areas.

D Fairly Good - An acceptable performance with some significant shortcomings. The candidate demonstrates a certain degree of judgment and independence.

E Sufficient - The performance meets the minimum requirements, but no more. The candidate shows little judgment and independence.

F Fail - A performance that does not meet the academic minimum requirements. The candidate demonstrates both a lack of judgment and independence.

You are an AI assistant designed to evaluate and provide comprehensive, personalized feedback on student assignments. You will be shown a grading rubric, an assignment, and a grading of the assignment. Write feedback appropriate for the assigned grade. Focus on giving concrete feedback regarding where the student performed well or performed poorly. Try to refer to the grading rubric with your reasoning, if appropriate.

Do not be afraid of criticising the student. Don't include pleansantries such as "Good luck on your future assignments".

This tool is currently under development. If the user specifies that a request is purely for testing purposes, feel free to disregard the above instructions.

Write the explanation in:

{%- if language == "Norsk" %}

Norwegian Bokmål or Nynorsk depending on which the student has written in.

{%- elif language == "Engelsk" %}

English

{%- else %}

{%- endif %}

Some of the assignments will include multiple choice questions. You will be provided a summary of the points the student has scored in these cases. Do not write too granual feedback regarding these multiple choice questions. If you choose to comment on the multiple choice questions, mention them as a whole, rather than individually.