Using AI under Norway's Personal Data Act

Norway does not yet have an AI Act in force. The EU AI Act has not been incorporated into the EEA Agreement, and the government plans to put an amended bill to consultation in autumn 2026 before presenting it to the Storting in spring 2027, as set out on the Norway AI Act page. In the meantime, the rules that bind most AI use in Norway are data protection rules. This guide covers what Datatilsynet expects when an organisation adopts AI tools, as of October 2026.

The legal base

The Personal Data Act of 15 June 2018, in force since 20 July 2018, makes the GDPR Norwegian law. Every principle of the GDPR applies to AI that processes personal data. That means lawfulness, purpose limitation, data minimisation, accuracy, storage limitation, security and accountability, together with the rights of the people whose data is used. Where an AI system makes decisions about people without meaningful human involvement, GDPR Article 22 applies.

Norwegian employers also have to consider the regulation on employers’ access to e-mail and other electronically stored material, known as the e-mail regulation, which prohibits monitoring of employees’ use of electronic equipment except in narrow cases.

What Datatilsynet has said

Datatilsynet has run a regulatory sandbox for responsible AI since 2020 and made it a permanent offer after an evaluation in 2023. Its sandbox reports are the most practical statement of how it reads the law for AI. The most widely relevant is Copilot through the lens of data protection, its November 2024 report with the Norwegian University of Science and Technology (NTNU) on Microsoft 365 Copilot. Its findings apply to most generative AI tools that work across an organisation’s own data.

The platform comes first. A tool built on a cloud platform depends on that platform’s security and data protection assessments being done, and on the organisation having the expertise to manage the provider over time. Responsibility for the data used stays with the organisation.

Get your own house in order. An AI assistant sees what its user can see. Weak access management and poor control of personal data are made visible and amplified by it. Datatilsynet expects a review of information management, the record of processing activities and access control before rollout. NTNU itself concluded it was not yet ready to roll Copilot out across the university.

Decide what it is for. Identify the tasks the tool should and should not be used for. Datatilsynet singled out HR and personnel management as particularly high risk, and said tasks involving special category data should be carefully assessed or avoided. Information containing personal data should be classified and labelled, at least at document level.

Check the legal basis. For each candidate task, check whether using the tool changes the processing and whether the existing legal basis still covers it, including whether the processing is still necessary. New purposes need their own legal basis or a compatibility assessment.

Assess the impact. As a general rule a data protection impact assessment will be required for generative AI that processes personal data. It is needed for each processing operation or set of similar operations, and the measures it identifies must be in place before the tool is used. If the risk stays too high, Datatilsynet suggests not using the tool for that processing, or seeking prior consultation.

Train the people using it. The organisation is responsible for making sure users have enough knowledge and awareness to use the tool well and in a way that protects personal data.

Consider alternatives and go step by step. Lower-risk tools built for a specific purpose may meet the need better. Where a general assistant is used, Datatilsynet strongly recommends introducing it in small, controlled steps with selected roles, with a plan for monitoring the quality of what it produces.

Putting it into practice

Datatilsynet’s sequence translates into a short set of controls that work for any AI tool, not only Copilot.

• An inventory. Know which AI tools are in use, including those staff adopted on their own. A shadow AI review usually finds more than the procurement list shows.
• Approved purposes. Record for each tool what it may be used for, which roles may use it and which data it must not see.
• A DPIA per purpose. Build the impact assessment on the record of processing, purpose by purpose, using a consistent impact assessment method.
• A usage policy, enforced. An acceptable use policy tells staff what they may enter, and guardrails stop personal data going where the policy says it should not.
• Monitoring after rollout. Check output quality and how use spreads beyond the roles first approved.

Preparing for the AI Act

The same records carry forward. Once KI-loven is in force, each system will need an AI Act role and risk tier, deployers of high-risk systems will need human oversight and monitoring, and some deployers will need a fundamental rights impact assessment. Datatilsynet has said that assessment resembles a DPIA, so an organisation that has built its DPIAs on a clear description of each system and purpose has most of the material already.

Where Aicura fits

Aicura supports this work. Discovery finds the AI tools in use, and the AI register records each one with its approved purposes, owner and the personal data it touches. Assessments hold the impact work for each purpose, guardrails apply your usage rules to live traffic, and monitoring tracks behaviour after rollout. The Essentials plan adds assessment against Datatilsynet’s guidance on AI and the EDPB’s opinion on AI models. It supports the work and the decisions stay with you.