Sweden Blocks 2 AI Chatbots, Keeps Copilot [2026]

Sweden’s central government has stopped staff from using two well-known AI chatbots for everyday office work. According to TechRadar, the Government Offices (Regeringskansliet) began blocking the web versions of OpenAI’s ChatGPT and Anthropic’s Claude on staff work computers on October 26, 2026. Officials say the reason is information security. However, Sweden still allows staff to use Microsoft Copilot.

The change affects only certain ways of using these tools, but it sends a clear message. It does not affect the public, ban either company, or stop all government employees from accessing these chatbots. Instead, it separates AI tools covered by formal agreements from tools employees open in a browser and use without the same level of official control.

This matters beyond Stockholm. Our earlier report on workers using unapproved AI tools showed that employees often turn to chatbots their employers have not approved when workplace rules fail to keep up. Sweden’s response is to block these open services and direct staff toward tools that have been checked. This article explains what was blocked, what remains available, what officials have said and what the decision could mean for OpenAI, Anthropic, Microsoft and IT teams elsewhere.

What Sweden’s Government Offices Blocked on October 26

The Government Offices is the central administrative organisation that supports Sweden’s government cabinet. Its staff include civil servants, advisers and policy workers. When its IT department changes which websites people can access, the change affects everyday work across the organisation. On October 26, the new rule targeted the web versions of ChatGPT and Claude on staff computers.

Reports describe a restriction on government work devices, not a rule for the whole country. A civil servant using a personal laptop at home is not covered by this decision, and neither is the general public. This distinction is important. Headlines saying Sweden has banned two chatbots across the country make the action sound broader than it is. In fact, one government employer has blocked access to two services on its own network.

Sweden’s public broadcaster reported the decision under a headline that roughly translates as ChatGPT and Claude being prohibited at the Government Offices. The Sveriges Radio report is the main Swedish-language source for the story. English-language reports from TechRadar, Euronews and Notebookcheck describe a similar situation.

Why Information Security Drove the Decision

The main concern is what happens when a government employee enters text into a public chatbot. That text leaves the government’s controlled systems. Without a formal agreement explaining how the company stores, handles and protects the information, the government has limited contractual control over what happens to it next. TechRadar, Euronews and Notebookcheck all describe the lack of formal agreements covering data protection and processing as a reason for the restrictions.

This is about contracts and purchasing decisions as well as technology. A chatbot used through a free or personal account usually comes with terms written for ordinary consumers. A business or government version can come with a negotiated contract, rules for handling data and management tools that an IT team can check. Sweden has decided that the first type is not suitable for official work, while the second may be acceptable when the right safeguards are in place.

The official explanation points to two types of risk. The first concerns what employees enter into a chatbot, especially sensitive information. The second concerns what employees do with the answers the AI produces. For example, if someone copies an incorrect number into a government briefing, that is not just a quality problem. It could also affect a policy decision. These risks can arise with any provider, which helps explain why the restrictions cover two competing companies.

One important point remains unproven. Reports do not confirm that Swedish government data was used to train ChatGPT or Claude, and they do not confirm a particular data leak. The Government Offices described examples of risky use, not a proven security breach. Calling this a confirmed data-leak scandal would go beyond the available evidence.

What the Digitalisation Director Said

Mikael Wettercrantz, the digitalisation director at the Government Offices, has provided the clearest explanation of the decision. According to Digit, he wrote: “We have seen several examples of problematic use of open AI services.” This suggests that officials based their decision on examples of how people were using these services, rather than on a general fear of AI.

Wettercrantz explained that the problem had two parts. As reported by Digit, “The examples concern both the sensitivity of the information that has been entered and how the AI-generated responses have been handled.” In simple terms, some staff entered information that should have been handled more carefully, while some did not treat the AI’s answers with enough caution. The Government Offices has not published the number of incidents, so there is no reliable figure to report.

On the issue of contracts, Euronews reported that Wettercrantz linked the use of AI services without an agreement to “consequences for our information security.” This is the main legal and practical reason for the decision. He was not accusing OpenAI or Anthropic of breaking the law or behaving badly. Instead, he was saying that using an outside AI service without a signed agreement creates risks the government does not want to accept.

The wording is measured. Officials did not describe a panic, name a specific security incident or call for a wider crackdown. The director described a workplace management problem and a way to address it. This is why the decision is best understood as an IT policy change.

Why Microsoft Copilot Stayed on the Approved List

One of the most important details is what Sweden chose not to block. Microsoft Copilot remained an approved AI tool after a security review. This makes the decision look like a difference in how the providers meet government requirements, even if that was not the intended message.

Reports say the Government Offices strengthened Copilot’s security settings after reviewing it. They also say a security agreement between Microsoft and the Swedish government covers the tool. These details do not prove that Copilot produces better answers than its competitors. They describe the steps taken to approve it: review the tool, adjust its settings, establish a contract and then allow its use. According to reports, the browser versions of ChatGPT and Claude did not go through the same process for the Government Offices.

Microsoft has an advantage because many government departments already use its systems for staff accounts, email and documents. Their legal teams may already have agreements with the company. Adding an AI assistant to these systems can build on an existing business relationship. Bringing in a new provider often requires a separate purchasing process, a new security review and a new agreement about how data will be handled. In cautious public organisations, an existing provider often has an easier path to approval.

The exact version and settings of Copilot used inside the Swedish government have not been fully disclosed. It is also unclear whether the Government Offices has enabled any Anthropic models that Microsoft can make available through Copilot. Claims about this internal setup should be treated as unverified until Swedish officials provide more details.

The Special AI-Assistant Platform Keeps Some Access Open

The restrictions do not completely cut off access to ChatGPT and Claude. Some employees can still use both services through a special AI-assistant platform developed for the Government Offices. This is an important difference between controlling access and banning the tools entirely.

How a government-built platform changes the risk

When a government builds or pays for its own system to connect to commercial AI models, it can decide what information is recorded, which employees can use it and what agreements apply to the AI provider. The service may feel much the same to employees even though the risks are different. Using a chatbot directly through its public website is like entering through an open door. Using it through a managed platform adds checks, records and contractual rules.

Reports link the platform to Intric, a company based in Stockholm. However, the Government Offices has not said how many employees can use it. Without that number, outsiders cannot tell how many staff have lost access to ChatGPT and Claude or how many have simply moved to a managed service. The platform suggests that the government wants to control how employees use AI rather than stop them from using it altogether.

What the platform tells us about demand

Governments rarely build internal AI systems for tools that nobody wants to use. Sweden’s decision to invest in a dedicated assistant platform suggests that employees had a real and continuing interest in generative AI. The block removes the uncontrolled way of accessing the tools, but it does not remove the demand for them. Employees may face some confusion and delays while they learn which tasks they can do with each approved tool.

Other Swedish Agencies Are Tightening Too

The Government Offices is not the only Swedish public organisation taking action. According to Sveriges Radio, the Swedish Tax Agency and the National Board of Institutional Care have also introduced restrictions on ChatGPT and Claude. Reports do not explain exactly what these restrictions involve, so it is not possible to say whether they are the same as the Government Offices’ block.

These agencies have very different jobs, but both handle sensitive information. The Tax Agency deals with financial records across the country. The National Board of Institutional Care handles some of the most private personal information held by public authorities. Both have good reasons to consider where information goes when an employee enters it into an AI chatbot. When organisations with such different responsibilities take similar precautions, it may point to a wider change in government policy rather than a one-off response.

Similar concerns are appearing in private companies. We reported on five major US banks blocking third-party AI agents. Their reasoning is similar to Sweden’s: organisations in heavily regulated industries want AI services they can control through enforceable agreements, rather than whatever tool an employee happens to find.

Open Chatbot vs Approved Tool: How the Access Routes Compare

The Swedish decision is easiest to understand by comparing three ways of accessing similar technology. The table below shows how published reports describe each option. It includes reported information without guessing about details the Government Offices has not shared.

The main difference is not how intelligent a model is, how well known its brand is or where it comes from. It is whether the organisation has suitable agreements and controls in place. A well-known model without a formal agreement lost access to government work computers, while a tool with a reported agreement and a completed review remained approved. For AI providers, this shows how important contracts and security checks are when trying to win government customers.

For people comparing models mainly by their abilities, our comparison of GPT-6.1 Sol, Sonnet 5.5 and Gemini 4 Argon covers pricing and performance. Sweden’s decision shows why an organisation may choose a different tool from the one that performs best in a test.

What Is Confirmed and What Is Still Unproven

Fast-moving news stories can attract claims that go beyond the facts. Separating confirmed information from unproven claims helps readers understand what happened and avoids unfairly blaming the companies involved.

The unconfirmed claims are especially important. None of the sources used for this article reports a confirmed data breach, a financial loss or a total number of affected employees. If a social media post or news aggregator claims otherwise, treat it as a rumour unless the Government Offices or a named news outlet supports the claim.

Historical Context: From Open Experiments to Contract-First AI

Public AI chatbots often entered workplaces without formal approval. Employees tried them using personal accounts, found them useful and began using them for everyday tasks before IT departments created rules. Many organisations initially allowed this to continue simply because nobody had set clear limits. Sweden’s decision shows a shift from ignoring the issue to creating an official policy.

The same pattern often appears across industries. A new tool spreads informally, concerns or incidents emerge, security teams raise questions, and leaders must decide whether to block the tool or bring it under a formal agreement. Blocking access is often the first response, while the longer-term solution is to offer a managed alternative. Sweden has taken both steps at once: it blocked open access to two services, provided a controlled platform and approved an enterprise product.

International politics also forms part of the wider background. Our report on the WAICO and Pax Silica AI blocs explains how countries face growing pressure to choose which AI systems and infrastructure they depend on. Sweden’s decision is much narrower and is focused on information security. However, it fits a wider trend in which governments treat AI services as important dependencies rather than just tools for getting work done.

Market Impact for OpenAI, Anthropic and Microsoft

The financial impact of blocking chatbot websites on one government employer’s computers is likely to be limited. Neither OpenAI nor Anthropic has reported a financial effect from the decision, and there is no reliable figure to give. The main effects concern reputation and future business opportunities, and each company is affected differently.

OpenAI and Anthropic

Both companies have lost direct access through their public websites at a visible European government organisation. However, their models remain accessible to some staff through the government’s special platform, suggesting that there is still interest in using them. The lesson is that consumer-style access is not enough to win many government customers. Providers need to offer suitable contracts, clear promises about data handling and management tools that cautious buyers can approve.

The timing is particularly important for Anthropic. The company is going through an initial public offering (IPO) process, which we covered in our report on Anthropic’s S-1 filing. Investors will want to understand how the company sells its products to customers with strict rules. A European government choosing a competitor’s approved product is one detail analysts may consider. It is a small piece of evidence, but it comes at a time when the company is facing increased attention.

Microsoft

Microsoft benefits most directly because Copilot remains approved while its two well-known rivals are blocked from direct browser access. This advantage comes from its existing business relationships, agreements and completed security review, not from proof that Copilot is better than the other tools. Even so, purchasing teams in other governments may look at Sweden’s decision and ask whether they should carry out similar reviews.

Usage figures help explain the wider problem. Our report on usage shares for Claude, ChatGPT, Gemini and Grok looks at how many people use these chatbots. Their popularity makes it difficult for organisations to prevent employees from using them without approval. Tools people become familiar with at home can create pressure to use them at work.

What IT and Security Teams Can Take From Sweden

Any organisation that handles sensitive information can learn from Sweden’s approach. It has four main parts, and none requires a government-sized budget.

• Block uncontrolled access. Public chatbots opened in a browser on work computers are difficult to manage, so they can be the first services to restrict.
• Approve an alternative that has been checked. Employees need another way to do their work. Sweden kept Copilot available after a security review and strengthened its settings.
• Provide a managed platform for special needs. Some teams need access to particular AI models. A controlled platform can provide that access without leaving the service open to everyone.
• Set clear rules about what information employees may enter. The Government Offices raised concerns about both sensitive information entered into chatbots and the way their answers were used. Workplace rules should cover both.

Organisations that cannot yet afford an AI service backed by a formal contract still have options. Our guide to private AI chatbots and their privacy models explains how different services handle information. This can help organisations write rules for acceptable use. Developers face similar questions when choosing coding assistants. Our Cursor vs VS Code network audit explains how to check what information an AI-enabled code editor sends outside the computer.

Our AI and machine learning section covers related developments, including new models, prices and the use of AI in businesses.

Predictions: Where Government AI Policy Goes Next

The following points are our expectations based on the pattern of Sweden’s decision. They are predictions, not confirmed facts, and they do not come from the Swedish government.

1. More public organisations will introduce limited blocks. Sweden’s tax and care agencies have already introduced restrictions. Other European ministries and agencies may review open web chatbots and reach similar conclusions.
2. Enterprise contracts will become essential. OpenAI and Anthropic may put more effort into agreements designed for government customers, because access through public websites alone is not enough to win government business.
3. Managed platforms will become more common. Governments that want particular AI models but do not want employees accessing them through public websites may prefer internal platforms or systems built for them, as Sweden has done.
4. Pressure to share more information will increase. Journalists and members of parliament may ask the Government Offices to publish the number of incidents and the number of people using its special platform.
5. Policies will focus more on approved tools than banned tools. Organisations may find it more useful to publish a list of approved services than simply list what employees cannot use, because demand for AI tools is unlikely to disappear.

Frequently Asked Questions

Did Sweden ban ChatGPT and Claude nationwide?

No. Reports describe a block on the browser versions of both services on Government Offices staff computers. This is not a nationwide ban, and the public is not affected.

When did the block start?

According to TechRadar and other news outlets, the Government Offices began blocking the services on October 26, 2026.

Why did the Swedish government block ChatGPT and Claude?

Officials say the reason is information security. They raised concerns about using AI services without formal agreements covering data protection and processing, as well as examples of problematic use of public AI services.

Is Microsoft Copilot also blocked?

No. Reports say Microsoft Copilot remains an approved AI tool after a security review.

Can Swedish government staff still use ChatGPT or Claude at all?

Some employees can. Reports say they can access both services through a special AI-assistant platform developed for the Government Offices. Officials have not disclosed how many employees can use it.

Was Swedish government data used to train ChatGPT or Claude?

This has not been confirmed. None of the sources used for this article reports that sensitive Swedish government data was used to train either model. The Government Offices has not published figures for incidents or data leaks.

Which other Swedish agencies have restricted these chatbots?

According to Sveriges Radio, the Swedish Tax Agency and the National Board of Institutional Care have also introduced restrictions on ChatGPT and Claude.

What should a company learn from Sweden’s decision?

Companies can block unmanaged chatbots on work devices, offer an alternative that has passed a security review and has a suitable contract, provide a managed platform for special cases, and set clear rules about what information employees may enter into AI tools.