Sweden Restricts ChatGPT and Claude While Retaining Microsoft Copilot for Government Staff

Table of Contents

• You might want to know
• Sweden’s Government AI Policy
• Key Insights Table
• Afterwards...

You might want to know

• Why is Sweden blocking ChatGPT and Claude on government staff computers while continuing to allow Microsoft Copilot?
• Could Claude models still be available to Swedish civil servants through Copilot or an internal AI platform?

Sweden’s Government AI Policy

Sweden’s Government Offices, the administrative body that supports the prime minister and cabinet, plans to block ChatGPT and Claude on staff computers starting Monday, October 26. Civil servants are being directed to Microsoft Copilot, which will remain the only approved AI tool for general use on those devices. The decision reflects a growing concern among public institutions about how information entered into consumer AI services might be handled, especially when the services are used without a formal agreement with the provider.

The policy was reported by Sveriges Radio, Sweden’s national, publicly funded radio broadcaster. Mikael Wettercrantz, the Government Offices’ director of digitalisation, explained the rationale in a post on the staff intranet. He warned that using AI services without a formal contract can have “consequences for our information security.” The concern is not limited to the technology itself; it also involves what employees enter into a service and how they use or distribute the responses it produces.

Wettercrantz said officials had seen several examples of problematic use of open AI services. Those cases involved both the sensitivity of information submitted to the tools and the way AI-generated responses were handled. The statement does not identify the specific incidents, the employees involved, or whether any particular case resulted in a confirmed data breach. Instead, it describes a broader risk-management concern that has prompted the Government Offices to distinguish between consumer services and tools reviewed under formal arrangements.

That distinction matters because consumer AI tools may transmit user queries to servers operated by external companies. The source article notes that U.S. law allows authorities to compel companies to hand over data when requested. This does not establish that data entered by Swedish employees has been obtained by authorities, nor does it describe a particular request. It does, however, help explain why government organizations may be cautious about sending sensitive material to services that operate outside a defined public-sector contract.

A formal security agreement can set out what a provider may and may not do with information submitted by government employees. It can establish limits and responsibilities for handling inputs, which is different from relying on the terms of a general consumer service. In this context, the Government Offices’ decision is framed as a question of governance and security controls, rather than a judgment that all AI tools are inherently unsuitable for public work.

Microsoft Copilot remains available because the Government Offices reviewed it and raised its security settings after that review. The source says no equivalent review and security-setting process was carried out for ChatGPT or Claude. The difference is therefore based on the Government Offices’ assessment and configuration of the products, not simply on whether the tools use generative AI. The announcement does not provide a detailed list of the review criteria, the specific security settings changed, or the conditions under which staff may use Copilot.

There is also a limited exception to the apparent ban. Some employees can still access ChatGPT and Claude through an AI platform built internally by the Government Offices. The press office has not said how many employees have this access or what restrictions apply. This detail means that the policy is not necessarily a complete prohibition across every government-operated environment. Instead, access may depend on the platform, its controls, and the circumstances of use, although the available information does not spell out those distinctions.

Sweden’s decision follows similar actions at other public agencies. The Swedish Tax Agency and the National Board of Institutional Care have already imposed comparable bans on consumer AI tools. The Enforcement Authority, which is responsible for debt collection and enforcing court orders, is preparing a decision of its own. Separately, an internal audit at the Transport Administration found roughly 10 AI services being used in what it described as a “widespread and risky way.” The account does not identify all of those services or provide further details about how the audit assessed the risks.

The policy also fits into a wider European debate over the use of AI in government. In February, the European Parliament disabled AI features across all its work devices, including ChatGPT, Claude, and Copilot. Its stated reason was that it could not guarantee the security of data processed by those tools. The Parliament’s approach was broader than Sweden’s: it turned off features across the three named products, whereas the Swedish Government Offices has retained Copilot after conducting a dedicated review and changing its settings.

Public-sector concerns have been developing for several years. The source traces the European debate back to Italy’s ChatGPT ban in 2023. Sweden is now taking a narrower approach than the European Parliament, approving one U.S. technology company’s product while excluding two others from staff computers. This comparison illustrates that institutions may reach different conclusions based on their own reviews, contracts, technical controls, and assessments of what information their employees handle.

A further complication is the relationship between Copilot and Claude models. Microsoft incorporated Claude models into Copilot in early 2026. In the European Union, where Sweden is located, Anthropic’s models are off by default and require administrators to deliberately enable them. The Government Offices has not confirmed whether Claude is enabled in its Copilot instance. As a result, the distinction between the prohibited product and the approved one may be less straightforward than the initial announcement suggests.

This uncertainty does not establish that Swedish government employees are using Claude through Copilot. It means only that the Government Offices has not confirmed the relevant configuration. The practical implications depend on how its Copilot environment has been set up and what administrators have chosen to activate. That missing information is important because the announced policy draws a clear line between named services, while the underlying technology and product integrations may create less obvious boundaries.

For now, the Government Offices’ stated position is that Copilot is the approved AI tool for staff computers following a review and additional security settings, while ChatGPT and Claude are blocked in that environment from October 26. Some access to the blocked tools remains possible through an internally built platform, but the scope and restrictions have not been disclosed. The policy therefore combines a visible restriction with exceptions and technical details that remain unresolved.

Key Insights Table

Afterwards...

The next phase of Sweden’s policy will depend on how the Government Offices defines and enforces its approved-use rules. Greater clarity about the internal AI platform, Copilot’s configuration, and the review criteria could help staff understand which tools and workflows meet the organization’s security requirements. Other public agencies may also reach decisions of their own as they assess consumer AI use and formal provider agreements.

More broadly, the case shows why public-sector AI policies may need to consider not only product names but also contracts, settings, data flows, and integrations between models and platforms. Sweden’s current approach is narrower than the European Parliament’s device-wide restrictions, yet its unanswered questions show that a clear policy announcement does not always resolve every technical boundary. Future guidance and transparent configuration details will help determine how the government balances AI access with information security.