Sweden's Government Offices Block ChatGPT and Claude on Work Computers

Sweden’s Government Offices, the Regeringskansliet that serves the prime minister and central ministries, is blocking access to ChatGPT and Claude from staff work computers, with the restriction taking effect Monday, per Sveriges Radio and SVT. Employees are being directed to Microsoft Copilot, the only AI service approved for use after a review in which its security setup was upgraded. The Government Offices did not provide comment to either broadcaster, but the shape of the decision is familiar across Swedish agencies: Skatteverket and the National Board of Institutional Care already block similar tools, and the Swedish Enforcement Authority is reportedly considering the same step.

The approval model, spelled out

What makes this worth attention beyond Scandinavia is that Sweden is not doing what most organizations do, which is issue a policy memo and hope. The Regeringskansliet implemented a network-level block and paired it with a sanctioned alternative that cleared a security review. That is the complete procurement loop most enterprises still have half of: ban the unsanctioned tools, then actually provide one that passed review, so employees have somewhere legitimate to go. The contrast with the United States is instructive; the DHS order removing DeepSeek from federal civilian agencies targets a foreign model on supply-chain grounds, while Sweden’s block applies to the two most widely adopted Western assistants on data-residency grounds. The criteria differ, but the destination is the same: government work consolidates onto whatever vendor can sign the security paperwork, and right now that is Microsoft, whose Copilot benefits from exactly this kind of review everywhere it runs.

The irony, and the real trade-off

Swedish coverage inevitably notes that Prime Minister Ulf Kristersson said in 2025 that he uses AI tools “quite often,” reportedly even for governing counsel. The deeper trade-off is the one every CISO is negotiating: blocking frontier consumer tools trades capability gains for containment. Copilot inside a government tenant offers contractual data handling and audit trails that ChatGPT and Claude consumer plans do not, and Sweden judged that worth more than frontier-model quality. Whether that judgment ages well depends on how large the capability gap becomes; if frontier assistants pull further ahead, agencies will face pressure to run sanctioned versions of the blocked tools through enterprise agreements, which is precisely the Azure-hosted and VPC-deployed model Anthropic and OpenAI have been building out for exactly these buyers.

What to watch

Two threads matter for anyone managing AI access in a regulated organization. First, whether other Swedish agencies follow: Kronofogden’s deliberation is the immediate one, and a national-government-wide block would make Sweden a case study in the approve-one-vendor model. Second, whether Copilot’s approval survives contact with usage: government employees routed to a single tool tend to generate the edge cases (classified-adjacent material, foreign-language documents, bulk processing) that stress any deployment. This is also the quiet counterpoint to this month’s consumer agent launches: Claude Duet and Muse are competing to act on users’ behalf, while governments are still litigating whether staff may paste a document into a chatbot. The gap between those two paces is where the next several years of enterprise AI policy will be written.