Civil servants at Sweden’s Government Offices, the machinery that serves the prime minister and every ministry, found two familiar websites out of reach on Monday morning. From October 26, ChatGPT from OpenAI and Claude from Anthropic can no longer be opened in a web browser on official computers, a change first reported by public broadcaster Sveriges Radio last week.
The move is not a retreat from AI as such. Staff are being pointed to Microsoft’s Copilot, which the Government Offices says it has reviewed and tightened, and which is covered by a contract with Microsoft spelling out how official data is handled. A limited circle of employees keeps access to the OpenAI and Anthropic models, but only through Intric, a Stockholm firm whose assistant platform, it says, runs entirely on servers inside Sweden.
The trigger was an internal audit carried out this year at Trafikverket, the national transport agency, which found what officials called “problematic” use of chatbots for government work. The fear is a familiar one in IT departments: when sensitive text is pasted into a consumer AI service with no agreement behind it, nobody can say for sure whether it ends up stored, used for training or processed in a data centre outside Europe. Digital chief Mikael Wettercrantz said unapproved tools of this kind put the security of government information at risk.
Other Swedish agencies got there first. Skatteverket, the tax authority, and SiS, which runs youth and addiction care homes, already limit both chatbots, and the Enforcement Authority is weighing a similar rule. Sweden had already put limits on the Chinese service DeepSeek.
Neither OpenAI nor Anthropic has commented publicly. For other European administrations wrestling with staff who quietly use whatever AI tool is handy, Stockholm’s approach offers a middle path between a blanket ban and a free-for-all: keep AI, but only through doors that have been checked and signed for.
