Sweden’s central government administration has blocked direct web access to ChatGPT and Claude on work computers as of October 26, citing information-security concerns. The restriction applies to Regeringskansliet, the government offices supporting Sweden’s prime minister and ministries; it is not a nationwide ban on public use of the AI services.
| Service | Status on government work PCs | Access condition |
|---|---|---|
| ChatGPT | Direct web interface blocked | OpenAI models may be available through Intric |
| Claude | Direct web interface blocked | Anthropic models may be available through Intric |
| Microsoft Copilot | Approved | Allowed following an internal security review |
The policy focuses on how government information is handled rather than declaring OpenAI or Anthropic’s models unusable. Some staff can still access those models through Intric, an internal AI-assistant platform designed to provide a controlled environment.
The decision is not aimed at the AI model itself, but at the data path: who manages it, what contract is in place, which policies apply and who is accountable if a sensitive document reaches the wrong place.
Mikael Wettercrantz, Regeringskansliet’s director of digitalization, warned that use of AI services without formal agreements can create information-security consequences. The central concern is unsanctioned “shadow IT,” where an employee might paste a document into an external chatbot for summarization, rewriting or translation without sufficient control over where the content is processed.
Microsoft Copilot remains permitted after a security assessment and additional safeguards, although the government has not disclosed its technical configuration. Its approval does not eliminate AI-related data risks: Microsoft acknowledged a bug in February after Microsoft 365 Copilot Chat accessed and summarized emails marked confidential.
A wider public-sector shift
Other Swedish public bodies have already adopted similar limits, including the Swedish Tax Agency and Statens institutionsstyrelse, which operates certain care and detention facilities. The Swedish Enforcement Authority is preparing its own decision, while the Swedish Transport Administration identified risky use of multiple AI services in an internal audit.
The move reflects a broader government challenge: generative AI can speed up drafting and analysis, but the risk changes sharply when tools receive internal correspondence, personal data or policy documents. Sweden’s approach keeps approved AI tools in use while attempting to place their data handling within a controlled and accountable framework.
