Denmark - Responsible Generative AI Guidelines

Overview

The Digitaliseringsstyrelsen guides published in January 2024 provide short, practical, organisational guidance for responsible adoption of generative AI by public authorities and private businesses in Denmark. They present clear, managerial-first advice: senior management should own the policy choices, organisations should classify how AI will be used (general vs. specialised), create tailored rules for staff, and set up the organisational capacity to test, validate and monitor AI use. The guides are explicitly designed to be updated as the technology and the legal framework evolve and to serve as an accessible starting point for internal governance rather than as technical standards or legally binding regulation. For the official announcement and to download the guides, see Digitaliseringsstyrelsen news release (09.01.2024). ([digst.dk](https://digst.dk/nyheder/nyhedsarkiv/2024/januar/nye-guides-til-ansvarlig-anvendelse-af-generativ-kunstig-intelligens/?utm_source=openai))

Definitions

The guides define key terminology with concise descriptions to help non-technical leaders: "generative AI" (models that produce text, images, audio, video or code from prompts), "foundation models" (large models trained on broad data that can be adapted to many tasks), "prompts" (user inputs to generate outputs), and distinctions between "open source" and "closed source" models and between cloud services and locally deployed models. The documents emphasize that generative AI is statistical in nature — it predicts likely outputs rather than reveals facts — which underlies risks such as hallucinations and outdated or incorrect outputs. These definitions are aligned with the explanatory sections on the guides' web pages and downloadable PDFs. ([digst.dk](https://digst.dk/kunstig-intelligens/guides-til-brug-af-kunstig-intelligens/guide-til-virksomheder/))

Governance and Institutional Framework

Digitaliseringsstyrelsen places governance responsibility at the level of senior management: organisations should treat responsible AI as a leadership task, ensuring allocation of resources and clear roles for policy, procurement, cybersecurity, legal review and operational rollout. The guides recommend a three-step governance workflow: 1) decide on an approach to AI use (scope and modality), 2) draft and publish organisational rules and guidelines adapted to that approach, and 3) build organisational capacity (training, designated owners, processes for testing and validation). For public authorities, the guides underline additional administrative obligations such as confidentiality and procedural fairness; for private companies they highlight contractual and IP considerations. The guidance also points readers to Digitaliseringsstyrelsen's other materials and preparatory resources as Denmark aligns with the EU AI Act and national oversight planning (guides index). ([digst.dk](https://digst.dk/kunstig-intelligens/guides-til-brug-af-kunstig-intelligens/guide-til-virksomheder/))

Key Focus Areas

The guides concentrate on risk-aware decision-making across multiple domains: (a) Accuracy & reliability — recognising hallucinations and designing human validation and fact-checking processes; (b) Bias & fairness — screening for discriminatory outputs and ensuring oversight where decisions affect people; (c) Data protection & confidentiality — avoiding input of personal data or trade secrets to third-party, free cloud models and carrying out DPIA-like considerations; (d) Security — preventing leakage of sensitive data and ensuring secure integration with enterprise systems; (e) Vendor & contract risk — obtaining contractual assurances about data use, retention and model training rights; and (f) Organisational readiness — staff training, logging and documentation. Practical mitigations are provided for each focus area — e.g. human in the loop, limited prompt scopes, using paid/enterprise services with stricter contracts, running models locally for sensitive use-cases, and maintaining audit trails of outputs and decisions. The approach is intentionally pragmatic: short guidance items and checklists that managers can adapt to their organisational context. ([digst.dk](https://digst.dk/media/kz3dgblm/guide-til-virksomheder-om-generativ-kunstig-intelligens.pdf))

Implementation Framework

The guides give a stepwise implementation path: start with an inventory and classification of use-cases (who, what, sensitivity), select an approach (general availability for staff or project-specific deployments), draft retningslinjer (internal rules) addressing permitted uses and prohibitions (e.g., do not enter personal data into public free services), and create organisational arrangements (designated owners, resources for testing, training and monitoring). They recommend concrete contractual checks when procuring cloud or SaaS AI services, including clauses on provider rights to use prompts and outputs, data retention, and model training. For high-sensitivity or regulated contexts the guides point to the option of local deployment or use of enterprise-grade offerings with stronger data protections. Implementation includes templates for communication to staff, training content, and a cadence for rule reviews. ([digst.dk](https://digst.dk/media/kz3dgblm/guide-til-virksomheder-om-generativ-kunstig-intelligens.pdf))

Monitoring and Evaluation

Monitoring focuses on tracking both technical performance (error rates, drift, hallucination incidents) and governance metrics (policy adherence, incident reports, vendor contract compliance). The guides recommend logging usage where feasible, maintaining records of decisions to rely on AI outputs, periodic review of model outputs for bias or inaccuracy, and mechanisms for users to report problems. For public bodies, continued review is also recommended to ensure compliance with administrative law obligations. Digitaliseringsstyrelsen indicates the guides will be updated based on experience and changes in legal requirements, and organisations are advised to treat monitoring as a continuous learning process rather than a one-off activity. ([digst.dk](https://digst.dk/media/0x5ptreb/guide-til-offentlige-myndigheder-om-ansvarlig-anvendelse-af-generativ-kunstig-intelligens.pdf))

Penalties, Liability, and Appeals

As non-binding guidance, the Digitaliseringsstyrelsen documents do not themselves introduce penalties. However the guides stress that failure to follow recommended practices can increase legal and regulatory exposure under existing frameworks — for example, GDPR obligations on personal data handling, procurement and contract law risks, sector-specific regulation (healthcare, finance), and administrative law for public authorities. Digitaliseringsstyrelsen therefore frames the guidance as risk mitigation: following the practices reduces the likelihood of regulatory enforcement, contractual liability and reputational harm. Organisations should therefore link their AI governance to their legal/compliance functions and maintain records that can support defence or remediation in case of complaints or audits. ([digst.dk](https://digst.dk/media/kz3dgblm/guide-til-virksomheder-om-generativ-kunstig-intelligens.pdf))

Relationship to Other Instruments

The guides are explicitly positioned as complementary to other Danish and EU instruments: they reference Denmark's national AI strategy and preparatory work for the EU AI Act, and should be used together with sector-specific rules, the GDPR, and national data protection guidance (Datatilsynet). Digitaliseringsstyrelsen signals that its guides are intended to help organisations prepare for future binding obligations under the EU AI Act and to operationalise existing obligations under privacy and administrative law. Where technical standards, conformity assessment or registration obligations appear under EU law, organisations will need to move beyond these guides to formal assessments and documentation. For more on related instruments see Digitaliseringsstyrelsen's AI pages and EU resources. ([digst.dk](https://digst.dk/kunstig-intelligens/guides-til-brug-af-kunstig-intelligens/guide-til-virksomheder/))

International Alignment

The approach in the guides follows commonly emphasised principles across EU and OECD guidance: risk-based oversight, human oversight for critical uses, protection of fundamental rights, data protection compliance and contractual due diligence with vendors. Digitaliseringsstyrelsen expressly notes alignment with the expected requirements of the EU AI Act and positions the guides as preparatory and pragmatic steps to help organisations comply with future binding rules. Organisations with international operations are advised to map these recommendations to other jurisdictions' rules and to use contractual and technical measures to harmonise compliance across borders. See Digitaliseringsstyrelsen's guidance index and the EU AI Act preparatory materials for further crosswalks. ([digst.dk](https://digst.dk/kunstig-intelligens/guides-til-brug-af-kunstig-intelligens/guide-til-virksomheder/))

Implementation Timeline

Milestone	Suggested timeframe	Notes
Initial awareness and inventory	0-3 months	Map existing uses and informal staff activity; quick wins: communications and immediate prohibitions (e.g., no personal data in public tools)
Draft internal rules & procurement checks	1-4 months	Establish permitted tools, vendor clause templates and human validation requirements
Pilot projects & security testing	2-6 months	Run supervised pilots, security / privacy tests and user training
Full organisational roll-out	6-12 months	Scale solutions where risk-mitigations are demonstrated; keep sensitive uses restricted
Ongoing monitoring & review	Continuous, quarterly reviews	Adapt to model updates, legal changes (notably EU AI Act timelines) and operational experience

Compliance Checklist

Item	Yes/No	Notes
Senior management assigned responsibility		Make role and owner explicit
Inventory of use-cases and sensitivity classification		Record: who, what, data types used
Internal rules for permitted/prohibited use		Include explicit ban on entering personal/confidential data to public tools
Vendor due diligence and contract clauses		Data retention, usage rights, training rights
Human-in-the-loop / validation for critical outputs		Define thresholds for manual review
Logging, documentation and record-keeping		Retention policy aligned with legal obligations
Training plan for staff		Regular sessions and update materials

Sources and References

Source	Type
Digitaliseringsstyrelsen: "Nye guides til ansvarlig anvendelse af generativ kunstig intelligens" (news release, 09.01.2024)	Primary Source
Digitaliseringsstyrelsen: Guide til virksomheder om ansvarlig anvendelse af generativ kunstig intelligens (PDF)	Primary Source
Digitaliseringsstyrelsen: Guide til offentlige myndigheder om ansvarlig anvendelse af generativ kunstig intelligens (PDF)	Primary Source