Since Monday, October 26, 2026, Swedish government staff no longer open ChatGPT or Claude in their browser. Sweden blocks ChatGPT and Claude on the workstations of Regeringskansliet (the Government Offices: the prime minister’s office and the ministries), according to SVT, which relays a report from the public radio broadcaster, revealed on October 23 by the Ekot program. Yet both models remain available to the same staff, through another door.
What Sweden blocks, and what stays open
The block targets the web interface, meaning the site you open in a browser. SVT (Gustav Motte and Mathilda Ståhl, article of October 23, updated on the 24th) quotes the instruction: « Webbläsargränssnittet för AI-tjänsterna ChatGPT och Claude kommer inte längre att nås » (the web interface of the AI services ChatGPT and Claude will no longer be reachable) from work computers. Staff are pointed to Copilot, Microsoft’s AI tool, which Regeringskansliet approved after a security review.
The detail that changes the picture is in the same article: « Inom ramen för Intric har medarbetare fortfarande tillgång till språkmodellerna ChatGPT och Claude » (within Intric, staff still have access to the ChatGPT and Claude language models). Intric is the administration’s internal platform, which serves the models inside a framework the state controls. The specialist site Feber also describes an in-house platform giving access to Claude and ChatGPT.
How many staff go through Intric, and with what limits? The crypto and tech outlet Decrypt writes that the press office did not say. Headlines like « Sweden blocks AI » therefore miss the point: only one way in is closed, and the models are not banned.
Why Sweden blocks ChatGPT: a contract, not a ban
Regeringskansliet did not detail its reasons to the media, says Feber. The known reasons come from an intranet message (the internal network) signed by Mikael Wettercrantz, director of digitalisation, which Decrypt reports secondhand. Using AI services without a formal contract can have « consequences for our information security ». He adds: « We have seen several examples of problematic use of open AI services ». The concerns cover how sensitive the information typed in is, and how the generated answers are then handled. We could not read that message ourselves.
The logic comes down to a contract. With a consumer offer, the request goes to the provider’s servers, and nothing governs what it does with it. With a signed agreement, or an internal platform, the administration sets the rules. Decrypt adds that US law lets authorities demand data from American companies: that is its own analysis, not a statement from the Swedish government.
Decrypt flags a gray area. Microsoft added Claude models to Copilot in early 2026, and the Swedish administration has not said whether that option is active in its Copilot. The same text notes that in the European Union, Anthropic’s models are off by default there and an administrator must turn them on.
Other Swedish agencies have already closed the door
According to SVT, the Tax Agency (Skatteverket) and the National Board of Institutional Care (Statens institutionsstyrelse, SiS) have introduced similar blocks. The Enforcement Authority (Kronofogden) is looking at the question. Decrypt adds that an internal audit at the Transport Administration found about ten AI services in use in a « widespread and risky » way. The idea that each employee picks their own assistant is therefore ending elsewhere than in Stockholm.
What this means for a French company
Sweden blocks ChatGPT on one precise point: the absence of a contract. A French small business can ask the same question without banning anything. Do your employees paste quotes, contracts or customer data into a personal account? An individual subscription gives you no say over what happens to what is typed, whereas a business offer sets in writing what the provider may do with it. Our article on ChatGPT’s settings in Chrome shows how much a consumer assistant sees of what is displayed in your tabs, and Claude Duet raises the same question for an agent that reads your emails while you are away.
On the French state side, the approach is an in-house platform. The AI assistant of the DINUM (the interministerial digital directorate) was opened on October 22, 2025 to 10,000 volunteer staff in eight ministries, for eight months, according to Acteurs Publics, which specifies that it is powered by the model of the French provider Mistral AI. The same outlet reports that Matignon had announced a contract with the provider for this chatbot in June 2025 (Acteurs Publics), and that a wider rollout was considered from June 2026 if the tool proved viable. We found no published assessment since. On the provider, see Mistral AI’s 3 billion euro raise.
0 Commentaires
Aucun commentaire pour le moment. Soyez le premier à commenter !