Sweden's Government Offices, the administrative body supporting the prime minister and cabinet, will block access to ChatGPT and Claude on staff computers starting October 26. The decision directs civil servants to use Microsoft Copilot as the sole approved artificial intelligence tool following a dedicated security review. Mikael Wettercrantz, the director of digitalisation for the Government Offices, cited information security concerns regarding the use of AI services without formal contracts. He noted that problematic usage patterns involved both the sensitivity of input data and the handling of generated responses.
The ban stems from risks associated with consumer AI tools that transmit user queries to external servers, where U.S. laws may compel data disclosure. In contrast, Copilot was vetted through a formal security agreement that defines provider obligations regarding government employee inputs. While ChatGPT and Claude are blocked on standard devices, some employees retain limited access via an internal AI platform built by the Government Offices, though specific restrictions and user counts remain undisclosed.
This move aligns with broader trends in Swedish public administration. The Tax Agency and the National Board of Institutional Care have already implemented similar bans on consumer AI tools, and the Enforcement Authority is expected to announce its own restrictions soon. An internal audit at the Transport Administration revealed widespread and risky usage of approximately ten different AI services. Internationally, the European Parliament disabled all AI features across work devices in February due to data security guarantees being unattainable. Sweden’s approach distinguishes itself by sanctioning one vendor while excluding others, despite Microsoft incorporating Claude models into Copilot in early 2026. The Government Offices has not confirmed whether Claude integration within their Copilot instance is active.
The Swedish Government Offices' decision highlights a critical divergence in how public institutions manage AI risk: prioritizing contractual control over technical capability. By blocking widely used tools like ChatGPT and Claude while approving Microsoft Copilot, the administration signals that legal frameworks and data sovereignty agreements are more significant than model performance or popularity. This creates a complex compliance landscape where the distinction between "banned" and "approved" software becomes blurred, particularly given that Copilot now integrates Anthropic's Claude models. The lack of transparency regarding whether these integrated features are enabled in the Swedish government's specific instance introduces potential shadow IT risks, undermining the stated goal of strict security containment.
From a Regulatory and Compliance perspective, this action reflects a tightening grip on generative AI within European public sectors, moving beyond blanket prohibitions toward selective vendor approval based on jurisdictional legal exposure. The reliance on U.S. law implications as a primary driver for banning consumer tools underscores the geopolitical friction surrounding data privacy. However, the existence of an internal workaround platform suggests that outright bans may drive innovation underground rather than eliminate it. Institutions must now balance the desire for secure, contractually bound AI solutions against the practical reality that employees will seek alternative methods to leverage advanced language models, potentially creating new vectors for data leakage if oversight mechanisms are insufficient.
