Sweden's Government Offices blocked employees from using browser-based ChatGPT and Claude on work computers starting Monday, 26 October, telling civil servants to rely on vetted alternatives after security reviews found sensitive material was being typed into chatbots that lack Swedish government data agreements.
The restriction covers the Rosenbad ministries and the broader Government Offices apparatus that supports the prime minister's work. Staff can still reach Microsoft's Copilot, which Stockholm has screened and covered under a security arrangement with the U.S. vendor. A limited group retains access to ChatGPT and Claude only through Intric, a Swedish platform that routes models without relying on foreign cloud contracts, according to briefings reported by Sveriges Radio and SVT.
Shadow IT on the public payroll
Mikael Wettercrantz, the Government Offices' director of digitalisation, told Ekot that AI services used without enterprise contracts can have "consequences for our information security." The move follows an internal audit at the Swedish Transport Administration earlier in 2026 that documented "problematic" use of roughly ten different AI tools, including cases where classified or sensitive inputs were pasted into consumer interfaces and where machine-generated answers were circulated without verification.
Ekot's reporting shows the pattern is wider than Rosenbad. The Swedish Tax Agency and the National Board of Institutional Care have imposed comparable blocks, while the Swedish Enforcement Authority is preparing a decision. DeepSeek had already been restricted at the Government Offices before this week's ChatGPT and Claude step.
Why Copilot stays while Claude goes
Stockholm's distinction is contractual, not ideological. Copilot passed a government security review and operates under negotiated protections for Swedish data. Consumer ChatGPT and Claude sessions, by contrast, leave officials without guarantees about retention, training use, or the jurisdiction where prompts are processed—an acute issue for EU public bodies after years of transatlantic cloud debate.
TechRadar and Euronews note that unregulated "shadow IT" can also complicate audits: when dozens of ministries experiment with different models, incident responders cannot easily trace what left the network or whether a leaked document was ever fed back into a vendor's training pipeline.
Europe watches the civil-service lab
Sweden is not banning generative AI from government work. It is centralizing how models are reached. That middle path—approve one enterprise stack, fence off browser toys—may become a template for other EU capitals as the European Commission enforces the Digital Services Act on large platforms and as national security agencies warn about foreign-hosted inference.
For vendors, the Swedish order is a market signal: public-sector revenue increasingly hinges on sovereign processing promises, not consumer-grade terms of service. For civil servants, Monday's block is immediate: if the chat window is not on the approved list, the help desk will say no.
Officials said Copilot settings were tightened after a fresh review, though they declined to publish technical details, citing security classification. More agencies are expected to follow the Transport Administration's audit trail in the coming months.