Sweden’s Government Offices Block ChatGPT and Claude. The Line Follows the Contract

Since 26 October, staff at the Swedish Government Offices can no longer open ChatGPT and Claude in the browser of their work computers. The Government Offices, the Regeringskansliet, are the organisation that serves the prime minister and the ministries. The public radio broadcaster Ekot broke the news on 23 October and public television SVT confirmed it with an email from the office. Staff are pointed to Microsoft Copilot. The tax agency (Skatteverket) and the national board that runs state care institutions (Statens institutionsstyrelse) already have similar blocks, and the agency that collects debts (Kronofogden) is considering one.

The stated reason is the lack of a contract. The office’s director of digitalisation, Mikael Wettercrantz, wrote on the intranet that using AI services the office has no agreement with can have consequences for information security, and he declined interviews. According to Bulletin, citing Ekot, the aim is to steer use towards approved tools, Copilot among them. Copilot’s security was tightened after an internal review, and the press office declined to give details because they concern security. The office had already restricted DeepSeek before these two.

I suspect the Swedish block follows the contract more than the model, and the detail that points that way is this: some staff will still reach ChatGPT and Claude through a platform for AI assistants that the office built itself. The same two models stay inside, through a door the office controls, and stay outside, through the door the vendor controls. What changes is who sees the documents an employee pastes in, and who answers if those documents leave the building. The line follows the contract, and the contract is where someone signs up to answer.

That said, an office that handles government documents has good reason to shut the public page of a service it has no agreement with, and none of the articles I read cites an incident behind the decision or says how many staff are affected.

What is missing is the criterion that let Copilot through. The review that cleared it is internal and the press office would not say more, so from outside nobody can tell a security decision from a purchasing decision: an approved list reads the same in both cases. It is the same mechanism as the pre-selected default: whoever writes the list decides for whoever uses it, and here the default is the list itself.

If your organisation is writing its own list of permitted tools, the useful question is what separates the blocked tool from the promoted one: the model, the contract, or a review nobody publishes. A list without public criteria shows only who signed.