Staff at Sweden's Government Offices are losing direct access to ChatGPT and Claude on their work computers. But they aren't being told to give up AI: Microsoft Copilot remains approved, and some employees will still be able to reach the blocked services through an internal platform.

The restrictions take effect Monday, October 26, according to Sveriges Radio's Ekot newsroom. They apply to the offices supporting the prime minister and cabinet, rather than every government worker in Sweden. Decrypt brought the decision to a wider audience in an October 24 report.

Two blocked websites, another way in

Digitalization director Mikael Wettercrantz told staff on the Government Offices' intranet that using AI services without agreements could have consequences for information security, according to the broadcaster's original report, published October 23. He declined an interview with the newsroom.

The aim is to get employees using tools the organization has approved. Officials also tightened Copilot's security after a review, the internal information showed. The press office declined to explain those changes, citing security concerns.

Then comes the exception. The press office said some staff would retain access to a special platform for AI assistants developed by the Government Offices themselves. That platform provides access to ChatGPT and Claude, even as the services become unavailable directly from employees' work computers.

That makes the route into an AI service part of the decision. The same names can appear on both sides of a workplace restriction when one way of accessing them has been approved and another hasn't. The broadcaster's report doesn't identify the platform's users or explain its technical safeguards, so the limits of that arrangement remain unclear.

An audit found problems with inputs and answers

The concern extends beyond Sweden's central government offices. An internal audit at the Swedish Transport Administration earlier this year found widespread, risky use of roughly 10 AI services, according to Ekot's reporting.

The audit described the problem this way, translated from Swedish:

We have seen several examples of problematic use of open AI services. The examples concern both the sensitivity of the information entered and how the AI-generated responses have been handled.

Those are separate workplace problems. Employees can expose sensitive material by putting it into a tool they haven't been cleared to use. They also need to check what comes back before treating an AI-generated answer as reliable. The audit raised both concerns, without publicly detailing the individual examples in the report.

Sweden's Tax Agency and National Board of Institutional Care have already introduced blocks of their own. The Enforcement Authority, which handles debt collection and enforcement of court orders, is preparing its position. The Government Offices had also previously restricted DeepSeek.

Copilot can include Claude too

There is another wrinkle in the approved alternative. Microsoft offers Anthropic models inside its own services, including Copilot. Its documentation says:

Customers within the EU Data Boundary and customers in the UK have Anthropic models disabled by default.

Administrators can choose to enable them. Sweden is in the EU, but the reporting doesn't establish whether its Government Offices have switched that option on. Copilot's ability to offer Claude therefore shouldn't be confused with confirmation that Swedish staff are using Claude through it.

For organizations, the account and contract also affect what happens to the information being entered. Microsoft's separate privacy guidance says for business customers, Copilot questions, answers and information accessed through Microsoft Graph aren't used to train the underlying AI models. It also says Copilot only shows work information that an individual user has permission to view.

Conversations can still be stored. Administrators can manage retention and search the saved interaction data using Microsoft's tools. The company also tells organizations using additional agents to check those agents' privacy statements and terms.