Sweden blocks ChatGPT and Claude on the work computers of its Government Offices from Monday, according to Swedish public broadcaster SVT, which cited Swedish Radio's news programme Ekot. Civil servants at the Government Offices (Regeringskansliet), the body that supports the prime minister and the ministries, will be pointed to Microsoft's Copilot instead. Copilot is the only AI assistant that has passed the office's own security review.
The decision is small in scale and large in signal. It is one of the clearest cases so far of a European central government choosing one AI supplier for everyday work and switching off the two best-known alternatives on its own network.
What Sweden's ChatGPT and Claude block covers
Based on the SVT report as relayed by Daily Northern and the German site Dr. Web, these are the facts that are on the record:
- Who: staff of the Government Offices of Sweden, on their work computers.
- What is blocked: OpenAI's ChatGPT and Anthropic's Claude.
- What is allowed: Microsoft Copilot. The Government Offices reviewed the service and increased its security settings before approving it.
- When: from Monday. The reports are dated Friday 23 October, which points to Monday 26 October.
- Elsewhere in the state: the Swedish Tax Agency (Skatteverket) and the National Board of Institutional Care (Statens institutionsstyrelse) have already blocked AI services. The Swedish Enforcement Authority (Kronofogden) is considering the question.
What is not on the record matters just as much. No official has been named as taking the decision, and no reason has been given for blocking ChatGPT and Claude specifically. SVT said it had asked the Government Offices for comment. We have not seen a published response.
Why a security review decides the winner
The one reason that has been given is about Copilot, not the tools that lost out: it was approved after a review and after its security was increased. That is how most public bodies now handle AI tools. Instead of judging which chatbot writes better, they ask a narrower set of questions:
- Where does the text that staff type in end up, and who can read it?
- Is the data kept inside a contract the organisation already controls?
- Can the IT department see and audit what is being used?
- Can access be switched off centrally if something goes wrong?
A tool that sits inside an existing Microsoft 365 agreement has an obvious head start on those questions. Most ministries already run their email, documents and identity systems on Microsoft. Adding Copilot means extending a contract and a set of controls that lawyers and security staff have already signed off. Bringing in ChatGPT Enterprise or Claude for Work would mean a new supplier, a new data agreement and a new review.
That does not mean ChatGPT or Claude failed a test. The reports do not say they were assessed at all. It means that, for now, they are not on the approved list, and the network will stop staff from using them anyway.
The political background in Stockholm
Use of AI at the top of the Swedish state has been a sensitive subject for more than a year. In August 2025 Prime Minister Ulf Kristersson drew criticism after saying in an interview that he sometimes used AI tools, including ChatGPT, to get a second opinion. Critics argued that a head of government should not be feeding questions about state business into commercial chatbots. His office said at the time that no sensitive information was involved.
A blanket block on the two most popular consumer-facing assistants removes that question for ordinary staff. Whatever officials do with AI on government machines will now run through one channel that the IT department controls.
How it compares with other governments
Sweden is not alone in drawing lines around which AI tools public servants may use, though the lines are drawn for different reasons.
- United States: the Department of Homeland Security ordered federal agencies to remove DeepSeek and tools built on it earlier this month, a national-security decision about a Chinese supplier. We covered it in DeepSeek Ban: DHS Orders Federal Agencies to Remove the Chinese AI.
- US federal buying: the General Services Administration's new AI clause sets contract terms for large language models bought by agencies, which shapes which suppliers can be used at all.
- Sweden: the target here is two American suppliers, OpenAI and Anthropic, and the beneficiary is a third, Microsoft. This looks less like a geopolitical decision and more like an IT governance one.
The common thread is that governments are moving from "staff may experiment" to "staff may use the approved tool". That shift usually comes with a single supplier, at least at first.
What is still unclear
Several questions will decide how much this matters beyond one government building:
- Personal devices. The block applies to work computers. The reports do not say whether staff may use other assistants on their own phones for work-related tasks, or whether that is banned by policy.
- Agencies. Sweden has a large number of independent agencies. The Government Offices decision does not automatically bind them, though the Tax Agency and the National Board of Institutional Care have already moved.
- Review criteria. If the security review that cleared Copilot is published, OpenAI and Anthropic would know what they need to show to be added. Without it, suppliers and the public cannot tell whether the rule is about data protection, cost or convenience.
- Lock-in. Choosing one supplier for AI at the centre of government makes the next decision harder. Procurement specialists will watch whether the arrangement is reviewed or simply renewed.
What it means for organisations watching
For ministries, regulators and large companies in Africa and elsewhere that are drafting their own AI usage rules, the Swedish move offers a simple template: approve tools one by one, after a security review, and block what has not been approved at the network level. It is easy to run and easy to explain.
It also has costs. Staff lose access to tools some of them may find more useful, and the organisation's AI options start to follow its existing software contracts rather than the quality of the tools. A written policy that explains why each tool was approved or refused, and how a supplier can apply, reduces both problems.
For OpenAI and Anthropic, the lesson is that winning public-sector users increasingly means passing government security reviews, not just offering an enterprise plan. Both companies sell versions of their products with data-protection terms aimed at exactly this kind of buyer. Whether those reach the Government Offices' approved list is now a procurement question, not a technology one.
We will update this story if the Government Offices publishes its reasons or the review behind the decision.
