Sweden's Government Offices has blocked web access to ChatGPT and Claude on staff computers, citing fears that civil servants were leaking sensitive files into commercial AI tools nobody had vetted.
The directive came from Mikael Wettercrantz, the agency's director of digitalization, who posted on the internal network warning that unregulated chatbots put national information security at risk. According to reporting from public broadcaster Sveriges Radio, the trigger was an audit at the Swedish Transport Administration that found employees routinely pasting confidential records into AI apps that had never been cleared for government use.
From now on, most civil servants will do their AI-assisted work in Microsoft Copilot instead. Broadcaster SVT reported that Sweden only approved Copilot after running it through a formal security review and locking down its privacy settings. The enterprise agreement reportedly guarantees that Microsoft cannot store government queries or use them to train future models, a protection consumer-facing tools like the free ChatGPT app or Claude's web interface don't offer by default.
A narrow exception, built at home
The ban isn't absolute. A select group of government employees can still reach ChatGPT and Claude, but only through a platform built by the Stockholm company Intric that runs entirely on servers inside Sweden. That setup keeps sensitive queries off foreign cloud infrastructure while still letting specific teams use the underlying models.
Other Swedish public bodies had already moved in the same direction. The Swedish Tax Agency has enforced a similar restriction, and the debt enforcement agency is reportedly drafting its own rules. Sweden had previously restricted the Chinese AI tool DeepSeek over concerns it didn't comply with the EU's GDPR privacy rules, so this isn't the country's first attempt to wall off government data from AI providers it can't fully audit.
The bigger worry behind all of it is what security teams call shadow IT: employees quietly adopting tools that were never approved, often because they're useful, free and one browser tab away. When those tools run on servers outside the country, and outside EU jurisdiction, a pasted spreadsheet or draft memo can become a compliance problem overnight. European data protection law carries steep penalties for exactly that kind of exposure.
Sweden isn't alone in drawing this line. Several United States federal agencies have placed similar restrictions on commercial chatbots on work devices, and the European Parliament went further earlier this year, disabling AI features entirely on staff devices when it couldn't confirm the data handling was safe.
What happens next will likely depend on how fast providers like OpenAI and Anthropic, the maker of Claude, can offer the kind of locked-down, audited enterprise terms that got Copilot over the line. Until then, expect more government offices across Europe to quietly do what Sweden just did: block the open web version and build, or buy, something they can actually vouch for.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.