Sweden's Government Offices Block ChatGPT and Claude for Staff

What the Government Offices did

Regeringskansliet, the Government Offices of Sweden, blocked access to ChatGPT and Claude on staff computers this week, first reported by Sveriges Radio's Ekot and confirmed by Expressen, which put the start of the block at this Monday. Staff are expected to use only AI services the offices have formally approved, with Microsoft Copilot the visible example, and Expressen reports that Copilot's security configuration has been strengthened at the offices at the same time.

The justification is contractual rather than technical. Mikael Wettercrantz, the offices' digitalization director, wrote on the intranet that using AI services without agreements in place can have consequences for information security. In other words, the test is not whether a chatbot is unsafe, but whether its vendor has signed the data-processing agreements a government requires before information can touch it.

The reason, and the internal escape hatch

Two details keep this from being a simple ban story. First, Regeringskansliet has built its own internal AI assistant platform, and part of the staff can use that platform to reach Claude and ChatGPT anyway, this time inside whatever agreements and logging the offices have set up. The block, in other words, is on unmanaged access, not on the models. Second, Copilot's approval is the same trade every enterprise has been offered: a consumer chatbot in exchange for a vendor who will sign the paperwork.

Ekot's survey found the pattern spreading across Swedish agencies: the Tax Agency, Skatteverket, and Statens institutionsstyrelse, the National Board of Institutional Care, have introduced similar restrictions. Sweden is therefore not regulating AI here at all; individual agencies are deciding, tool by tool, which services their employees may touch, and the deciding factor is procurement, not capability.

A wider pattern this fits into

Governments channeling staff toward sanctioned AI tools has become the standard move worldwide, and it runs in every direction at once. American legislators have proposed banning federal use of Chinese models outright, while Beijing now requires its own labs to register before shipping models abroad, as this blog covered earlier this week. The European Union's AI Act imposes its own transparency obligations on deployed systems. Sweden's block is the procurement-layer version of the same instinct: whichever models a government considers acceptable, the approved list will be short and the vendor list shorter.

What this means for AI at work

The interesting question for readers is why local models are absent from these conversations. A model running on government hardware with no external data flow would satisfy the security concern that motivated the block, and yet no agency in this story has gone that route. The reasons are mundane: the approved-tool procurement path already exists, someone has to maintain a local deployment, and Copilot is one checkbox on an existing Microsoft agreement.

That will not hold forever. The moment a government needs AI over sensitive material at scale, a sanctioned cloud chatbot and a locally hosted model start competing on the procurement's own terms, and the agencies building internal gateways, as Regeringskansliet has, are already halfway to that architecture. Watch which model the internal platform connects to next; that decision will say more about government AI policy than any ban.