Starting today, Sweden blocks ChatGPT and Claude on government work computers. The Government Offices, which runs the prime minister's office and every ministry, has cut staff access to the browser versions of both chatbots and is steering civil servants toward Microsoft Copilot instead. It is a government-level decision aimed at two American AI products, and it lands while US vendors face their own questions about agent containment and government data access.
Sveriges Radio's news programme Ekot first reported the plan on Friday, October 23. Notebookcheck reports that from Monday, October 26, the browser versions of both chatbots will no longer be reachable on staff computers. The change takes effect today, and the reasons behind it are more specific than a general unease about AI.
Why the Government Offices is cutting off ChatGPT and Claude
The reasoning comes from Mikael Wettercrantz, the Government Offices' director of digitalisation. In a post on the staff intranet, he wrote that using AI services without a formal contract can have "consequences for our information security," according to Madrobot's account of Ekot's reporting. That quote comes through a secondary report, and the original intranet post and Ekot's audio were not checked for this story.
The office points to two problems: missing contracts and several cases of problematic use. The available reporting does not describe those cases, so the public record on what went wrong is thin. The logic behind the decision is still clear. Without a contract, there are no defined data terms, and without defined data terms, it is hard to say where staff queries end up.
Decrypt's analysis draws that line sharply. Consumer AI tools used without formal agreements send user queries to external company servers, and US law allows authorities to compel those companies to hand over data on request. Decrypt makes that argument as its own analysis, not as a finding from the Swedish government. Still, it explains why contracts sit at the center of the decision.
Which tools are blocked, approved and still reachable
- Blocked: the browser versions of ChatGPT from OpenAI and Claude from Anthropic on staff work computers, starting October 26, 2026.
- Approved: Microsoft Copilot, which the office reviewed before raising its security settings. Crypto Briefing describes it as the only AI service cleared after a security review.
- Still reachable: some staff keep the ChatGPT and Claude models through Intric, a Swedish AI platform the offices procured in July 2026, according to CellCog's summary.
- Outside the block: AI as a category. The order covers specific services on work computers, not AI in general.
The press office told Ekot that certain employees can still reach ChatGPT and Claude through an AI-assistant platform the Government Offices has built itself. It did not say who those employees are or how many there are. That gap matters. A block that leaves an unknown number of staff with a route to the same models is hard to measure and harder to audit.
Copilot is not a clean substitute for Claude users either. Claude models were added to Copilot in early 2026, but BigGo Finance reports that the Government Offices has not said whether those models are enabled in its deployment. Pasquale Pillitteri's report speculates that Claude may return through Copilot. That remains speculation, not a confirmed plan.
Which other public bodies are following
| Body | Status as reported |
|---|---|
| Tax Agency | Similar ban already in place |
| National Board of Institutional Care (SiS) | Similar ban already in place |
| Enforcement Authority | Preparing a position |
| Transport Administration | Internal audit found roughly 10 AI services used in a "widespread and risky way" |
| European Parliament | Disabled ChatGPT, Claude and Copilot on its devices, per one report dated February; no second source confirmed the date |
Ekot reports that the same move has appeared at more agencies, and a third is weighing one. The Transport Administration finding approaches the problem from another direction. An internal audit there flagged the sheer volume of AI tools already in use, which points to a broader worry about unapproved services across the public sector.
The pattern is what makes Sweden's move notable. A single agency's ban would look like local housekeeping. Several bodies acting within months look more like a shared procurement view.
What the Swedish block does not do
The line between business and consumer accounts does much of the work here. The block applies to the computers staff use for work, and it is not a ban on AI. The office is telling employees which tools they may use and under which terms, not telling them to stop using AI.
Two details remain unclear: how many employees keep access through Intric, and what restrictions apply to them. Until the office publishes those figures, the block is best read as a policy on approved tools rather than a complete cut-off.
What the move means for US AI vendors
For OpenAI and Anthropic, this is a lost government channel in one European country. The bigger signal concerns who wins the contract. Microsoft gains a government advantage because Copilot cleared a review that the two standalone chatbots did not clear here. Public buyers are asking whether a product has a contract, a review and a data path they can defend before they ask which model is best.
The reporting contains no statement from OpenAI or Anthropic about the decision. Neither company has publicly answered the criticism in the coverage I reviewed, which leaves their side of the story untold for now.
Sweden's decision arrives while American attention is fixed on other problems. Our recent coverage of Anthropic's Claude pulling Georgia vote counts without human approval and of Cursor's computer use feature, which lets AI coding agents control a whole desktop shows where that attention is focused. Scrutiny centers on what an AI system can reach and what it does without a person approving the step. Sweden applies the same logic to procurement.
No US federal or state counterpart to the Swedish order appeared in the reporting. US federal workers and contractors are not directly affected by it. The practical relevance for American readers lies in how foreign public bodies rank US AI products, and what that ranking does to US vendors' revenue.
What US professionals should check now
The following steps are general guidance and do not come from the Swedish reporting.
- Check whether your employer has an enterprise agreement for an AI tool before you paste work documents into it.
- Ask IT which AI tools are approved for work, and whether consumer accounts are allowed on work devices.
- Use the company-approved version when one exists, and confirm its data retention and admin settings.
- Keep client, legal and personnel files out of personal chatbot accounts, even when the tool feels familiar.
The Swedish case suggests that account type often decides the question more than the brand of chatbot does.
What to watch next
Four developments matter in the coming weeks. Watch whether Claude returns through Copilot, which would show whether the restriction targets a vendor or a contract. Watch how many staff keep access through Intric and what limits apply. Watch whether the Enforcement Authority and the third agency follow the Tax Agency and SiS. And watch whether OpenAI or Anthropic respond with an offer aimed at Swedish public bodies.
If more agencies follow, the Swedish order will look less like a single procurement decision and more like a template for how European public bodies vet American AI tools. That template will shape which products reach government desks and which stay outside the door.
