Skip to content
PiiBlocker
Back to blog

Why Sweden's Government Blocked ChatGPT and Claude, and What Every Organisation Can Learn

PiiBlocker Team

ai policydata protectionchatgptclaudeshadow aigovernment

On Monday 26 October 2026, Sweden's Government Offices blocked ChatGPT and Claude on staff computers. The reason was not the AI. It was what staff had been typing into it: sensitive government information, entered into consumer AI services with no contract covering where that data goes. Sweden is the latest in a three-year run of similar bans, from Samsung to the US House of Representatives to the European Parliament. What makes Sweden worth studying is that it did not simply ban AI. It moved staff onto routes it could control, and that is the part every organisation can copy.

What happened in Sweden

The Government Offices, the administrative body that supports Sweden's prime minister and cabinet, blocked ChatGPT and Claude on staff computers from 26 October. The decision was reported by the national broadcaster Sveriges Radio.

Mikael Wettercrantz, the Government Offices' director of digitalisation, explained it to staff on the internal intranet. Using AI services without a formal contract, he wrote, can have "consequences for our information security", and "we have seen several examples of problematic use of open AI services."

It was not a sudden decision. A review earlier in 2026 of AI use at the Swedish Transport Administration found about 10 AI services in use in a "widespread and risky way".

Three details show this was a redirection, not a blanket ban:

  • Microsoft Copilot stays approved. It passed a security review, and the Government Offices raised its security settings afterwards.
  • Some staff keep ChatGPT and Claude. A limited number can still use them through an AI platform the Government Offices built internally.
  • Other agencies are doing the same. The Swedish Tax Agency and the National Board of Institutional Care have already restricted consumer AI tools, and the Enforcement Authority is preparing its own decision.

Why Sweden did it

Read the statements closely and the concern is specific. It is not that the AI gives wrong answers. It is about the data:

  1. No contract. Consumer AI services were used without a data-protection agreement setting out how information is stored and used.
  2. What was entered. Wettercrantz pointed to "the sensitivity of the information that has been entered".
  3. Where it goes. Data sent to a consumer AI service can be stored by the provider and, without an agreement in place, potentially used to train future models.
  4. How answers were handled. Responses from the tools were being used in government work without clear controls.

No breach was reported. This was a preventive decision: the Government Offices could not see or control what its people were sending, so it closed the uncontrolled route.

The twist: the approved tool can run Claude

The headline reads as "ChatGPT and Claude banned, Copilot approved". The reality is blurrier. Since early 2026, Microsoft Copilot has been able to run Anthropic's Claude models alongside its own. In the EU those models are off by default, and the Government Offices has not said whether they are enabled in its Copilot.

So the same underlying model can be blocked on one route and approved on another. That is the real lesson of the Swedish decision. What makes an AI tool safe to use at work is the agreement and the settings around it, not the brand name on the screen.

Sweden is not alone

Sweden's decision is the latest step in a clear pattern. It started with individual companies and has moved to parliaments and whole governments.

Date Who What they did Reason given
Oct 2026 Sweden's Government Offices Blocked ChatGPT and Claude on staff computers; Copilot approved "Several examples of problematic use"
Oct 2026 US Department of Homeland Security Ordered federal civilian agencies to remove DeepSeek products Order under the Federal Acquisition Supply Chain Security Act
Feb 2026 European Parliament Disabled built-in AI features on work devices IT could not guarantee the security of data uploaded to AI companies' servers
Feb 2025 Parliaments of Norway and Denmark Banned, or advised removing, DeepSeek National security authority warned of data-protection risks
Feb 2025 Australian Government Blocked DeepSeek on all government systems and devices Security directive (PSPF Direction 001-2025)
Jan 2025 Italy's data protection authority Blocked DeepSeek Protecting users' personal data
Mar 2024 US House of Representatives Blocked Microsoft Copilot on all House Windows devices "The threat of leaking House data" to unapproved cloud services
Jun 2023 US House of Representatives Banned the free ChatGPT; allowed the paid version with limits Data security
May 2023 Samsung Banned generative AI tools for staff Engineers had pasted internal source code into ChatGPT
Mar 2023 Italy's data protection authority Temporarily blocked ChatGPT nationwide Legal basis for processing, transparency, age checks

Note the US House entry. In 2024 it blocked Copilot, the same product Sweden approved this week, until a version built for government existed. Approval follows the agreement, not the product.

Two different reasons for a ban

The list above mixes two problems, and they need different answers.

"We don't trust this provider." The DeepSeek bans are about the company and the law it operates under. The concern is who could reach the data once it arrives, whatever that data is. Nothing a user does before pressing send changes that. For this kind of ban, blocking the tool is the right answer.

"Our people are pasting sensitive data into it." Samsung, the US House, the European Parliament and Sweden are a different case. The tools are mainstream. The worry is what staff type into consumer versions with no agreement in place. That is a behaviour and data problem, and a ban is only one way to address it, often not the best one.

Why bans on their own don't hold

"Shadow IT", staff using tools their organisation has not approved, names the weakness. People use these tools because they save real time on drafting, summarising and translating. Block them on work computers and many will use a personal phone instead, where IT sees nothing at all.

In a 2025 survey by the National Cybersecurity Alliance and CybSafe, 43% of respondents said they had already shared sensitive work information with AI tools without their employer's knowledge. A ban does not change that habit. It only moves it out of sight.

What Sweden got right, and how to copy it

Sweden's response has four parts. Any organisation can apply them at its own scale.

  1. Approve a tool under a data agreement. Use an enterprise plan with a data processing agreement, no training on your data, and admin control over retention. Sweden reviewed Copilot and tightened its settings before approving it.
  2. Give a controlled route to the tools people want. Sweden kept ChatGPT and Claude available to some staff through its own platform. Access moved; it was not simply removed.
  3. Set clear limits. Say which tools are approved, what data must never go into any of them, and who can use what.
  4. Remove personal data before it is sent. Even on approved tools, the safest prompt has no names, ID numbers or contact details in it. If the provider only ever receives [PERSON_A], there is far less to protect, retain or explain.

What that looks like in practice:

For the regulatory side, see our GDPR guide for DPOs. For how network-level controls compare with browser tools, see DLP vs. browser extensions for AI chatbots.

Where PiiBlocker fits

PiiBlocker handles step 4. It is a Chrome extension that detects personal data as people type or paste into ChatGPT, Claude, Gemini, Perplexity, Grok, DeepSeek and Kimi, and replaces it with placeholders before the message is sent. Detection runs in the browser, and the free version makes no network calls. For teams, IT can deploy it through Group Policy or a licence key and lock it on.

It is one layer, not a replacement for the others:

  • It does not cover desktop or mobile apps, other browsers, or uploaded files.
  • It does not address concerns about a provider itself. If you do not trust where a tool sends data, block the tool.
  • No detector catches everything, so clear rules on what goes into AI still matter.

For steps anyone can follow today, with or without an extension, see how to mask personal data before pasting into ChatGPT.

Frequently asked questions

Why did Sweden block ChatGPT and Claude? Sweden's Government Offices blocked them on staff computers from 26 October 2026 after seeing "several examples of problematic use", including sensitive information entered into AI services without a formal contract. Microsoft Copilot, which passed a security review, remains approved.

Can Swedish government staff still use ChatGPT or Claude at all? A limited number can, through an AI platform the Government Offices built internally. The block applies to the standard consumer services on staff computers.

Did ChatGPT or Claude leak Swedish government data? No breach was reported. The decision was preventive, based on the risk that sensitive information entered by staff was stored outside the government's control.

Is banning AI chatbots effective? It stops use on managed devices, but often pushes it to personal phones and accounts, where it cannot be seen. Approved tools, clear limits and removing personal data before sending tend to hold up better.

Does masking personal data solve the DeepSeek concerns? No. Those bans are about the provider and the law it operates under. Masking reduces the personal data a provider receives, but if you do not trust the provider, blocking it is the right call.

Sources

PiiBlocker is a free Chrome extension that masks personal data before it reaches AI chatbots. The free version runs in your browser and makes no network calls. Install from the Chrome Web Store.