Sweden’s Government Offices Block ChatGPT and Claude, Approve Copilot

Sweden’s Government Offices, known as Regeringskansliet, supports the prime minister and cabinet. Starting Monday, October 26, it will block staff from using the web versions of ChatGPT and Claude on their work computers. Microsoft Copilot will be the approved alternative. The Government Offices says it checked Copilot and strengthened its security settings before making it available to all employees. Sveriges Radio’s news service Ekot first reported the decision. SVT and Decrypt later confirmed it.

The decision is not so much a ban on two AI products as a decision about how government workers can access AI. SVT reports that staff can still use the ChatGPT and Claude language models through Intric, a platform used across government ministries. What is being blocked is the consumer version: a civil servant opening a chatbot in a web browser without using an agreement negotiated by the government. Sweden is joining a growing number of public organizations that are making this distinction. For business leaders, the case shows that AI rules depend on purchasing agreements and data protections, not just on which AI model is used.

What Changed, and Why

Decrypt reports that Mikael Wettercrantz, the Government Offices’ director of digitalisation, announced the change on the staff intranet. He said using AI services without a formal agreement can create information security risks. He also said the department had found several examples of problematic use of open AI services. These involved both the sensitivity of the information staff entered and the way they handled answers produced by AI.

His explanation points to two different risks. The first is data leakage. Information entered into a consumer AI service is processed on the provider’s servers under its standard terms, rather than under an agreement negotiated by the institution. The second risk is unreliable answers. Officials might treat AI-generated text as correct without checking it carefully. A formal agreement can help reduce the first risk by setting rules for how a provider can use information, how long it can keep data and where it can process that data. Reducing the second risk requires training, internal rules and clear responsibility for checking the results.

The Government Offices told SVT by email that staff would no longer be able to reach ChatGPT and Claude through their websites. Some other AI services would also be blocked, while staff would keep access to the ChatGPT and Claude models through Intric. The Government Offices’ press office did not tell Decrypt how many employees had access to these models or what limits applied to that access.

A Pattern Across the Swedish State

The Government Offices is not the only Swedish public body taking this step. SVT reports that Sweden’s Tax Agency and the National Board of Institutional Care (Statens institutionsstyrelse) have already introduced similar blocks. The Enforcement Authority is considering doing the same. Decrypt also reports that an internal audit at the Transport Administration found about 10 AI services in what it called “widespread and risky” use.

This audit finding may be one of the most important details in the story. In most organizations, employees started using generative AI before their employers officially approved it. Staff began using free tools to write drafts, summarize documents and translate text long before IT and legal teams set rules. Sweden’s public agencies are now trying to close that gap after the fact, as many companies are doing too.

European institutions have taken similar steps, although the rules differ. In February 2026, the European Parliament turned off built-in AI features on work tablets and smartphones used by lawmakers and staff. Politico and eWeek reported on an internal email in which the Parliament’s IT team said it could not guarantee the security of data that some features sent to cloud services. In 2023, Italy’s data-protection authority temporarily restricted ChatGPT. Each action focuses on data moving through systems without proper control, rather than rejecting AI itself.

Why Copilot Passed and Others Did Not

Some people describe Sweden’s decision as Europe choosing one American supplier over other American companies. That view misses the main reason for the decision. Microsoft is also a U.S. company and is subject to the same U.S. legal system, under which American authorities can sometimes require providers to hand over data. Copilot was not approved because it operates under a different national legal system. It was approved because the Government Offices already has an enterprise agreement with Microsoft, has reviewed Copilot as part of its existing Microsoft environment and controls how the tool is set up.

The differences between AI models are also less clear-cut than the announcement might suggest. Earlier this year, Microsoft added Anthropic’s Claude models to Microsoft 365 Copilot. According to UC Today, these models are turned off by default in the EU, EFTA countries and the U.K. because Anthropic is not covered by Microsoft’s EU Data Boundary commitments. Administrators must actively turn them on. Decrypt reports that the Government Offices has not said whether it has enabled them. This means the same AI model could be blocked when accessed through a web browser but allowed when used inside an approved business tool.

The same principle applies to Intric. Staff can continue using the ChatGPT and Claude models through this platform because it provides a controlled and approved way to access them. In practice, Sweden’s rule is that the agreement and security controls matter more than the model itself.

Why This Matters for Business Leaders

Many companies are in the same position Sweden’s government was in a year ago. Employees use consumer AI tools widely, but formal policies, approved platforms and legal protections for data have not kept up. Sweden’s experience points to several practical lessons.

First, a ban alone is not a strategy. If an organization blocks consumer AI tools without providing a useful approved alternative, employees may switch to personal devices, where the organization has less visibility and the risks may be greater. Sweden combined its restrictions with Copilot access for all employees and continued access to AI models through Intric.

Second, data agreements matter more than brand names. Company boards should know which AI services employees use, which services are covered by business agreements, whether information entered into them can be used to train AI models, how long data is kept and where it is processed. The country whose laws apply is important, but contract terms and system settings largely determine the actual risks.

Third, purchasing decisions now affect competition in the business AI market. Suppliers whose software is already widely used in an organization have an advantage when security reviews favor existing agreements. For AI developers, this makes it more valuable to distribute products through platforms such as Microsoft’s and to offer business agreements that meet the needs of European public organizations directly.

Fourth, organizations still need better rules for checking AI answers. Wettercrantz’s concerns about how staff handled AI-generated responses will not disappear just because employees switch to another tool. Organizations need clear rules for checking results, disclosing AI use and deciding who is responsible for work that AI has helped produce.

“Sweden has not rejected any AI model. It has rejected ungoverned access. That is the right lesson for every board: the competitive question in enterprise AI is no longer only whose model is best, but whose contract, controls and accountability can be trusted.” Proposed wording — pending speaker approval,” says Prof. Dr. Amarendra Bhushan Dhiraj, CEO and Editor-in-Chief of CEOWORLD Magazine, global economic strategist, and bestselling author of Between Olympus and the Ganges: How Greece and India Shaped the Modern Mind.

Executive Takeaways

What changed: Starting October 26, Sweden’s Government Offices will block web-browser access to ChatGPT and Claude, approve Microsoft Copilot and keep access to the models through an internal platform with controls.

Why it matters: European public organizations are separating AI use that is covered by approved agreements from consumer use without those controls. Regulated companies and their suppliers will increasingly be expected to meet this standard.

What leaders should do next: Check which AI tools employees actually use, make sure every approved tool has an enterprise agreement covering data use, and provide an approved alternative before restricting consumer services.

Sweden’s decision is unlikely to be the last of its kind. Over the next 12 to 24 months, important signs to watch will include whether more Swedish agencies introduce similar blocks, whether Sweden or the EU creates common rules for public-sector AI purchasing, and whether AI developers offer enterprise agreements and sovereign-cloud options that directly meet European governments’ needs. The broader lesson goes well beyond Stockholm: for institutions, competition in AI is shifting away from model quality alone and toward who can provide access that is secure, trusted and governed by clear contracts.