Sweden’s Government Offices Block ChatGPT and Claude. The Real Issue Is the Missing Contract

Since Monday, October 26, staff at the Government Offices of Sweden can no longer open ChatGPT or Claude on their work computers. The Regeringskansliet, which covers the Prime Minister’s Office and the government ministries, has blocked the browser versions of both chatbots and pointed employees to Microsoft Copilot, which it reviewed and then locked down further. The decision was first reported by public broadcaster Sveriges Radio.

The reason given was not that the models are dangerous. It was that the government has no agreement with the providers covering how its data is handled. That distinction is the most useful lesson in the story for any organization deciding what AI its staff may use.

What the Government Offices decided

Mikael Wettercrantz, the Government Offices’ director of digitalisation and chief information officer, declined an interview but wrote on the staff intranet that using AI services without an agreement could have “consequences for our information security.” (Quotes are translated from Swedish.) According to the reporting:

• ChatGPT and Claude are blocked in the browser on staff computers from October 26.
• Staff are directed to approved services, including Microsoft Copilot, whose security settings were tightened after an internal review. The press office declined to describe the changes, citing security.
• A limited group of employees keeps access to OpenAI and Anthropic models through an internal AI assistant platform the Government Offices built, which Swedish broadcaster SVT identified as Intric.
• The Government Offices had already restricted use of DeepSeek.

Other agencies are moving the same way. Sveriges Radio reported that the Swedish Tax Agency and the National Board of Institutional Care (SiS) have similar blocks, and the Swedish Enforcement Authority is preparing its own decision.

The audit that set the tone

The clearest evidence of the problem came from another agency. An internal audit at the Swedish Transport Administration earlier this year found widespread and risky use of around ten different AI services. “We have seen several examples of problematic use of open AI services,” the auditors wrote. “The examples concern both the sensitivity of the information entered and how the AI-generated answers were handled.”

That sentence describes two separate risks. The first is input: staff pasting sensitive or personal information into a service that may keep it, process it outside the EU or use it for training under consumer terms. The second is output: staff relying on generated answers without checking them. A tool ban addresses the first risk more directly than the second.

Why the contract is the control

Copilot is not inherently safer than ChatGPT or Claude, and the Government Offices did not claim it was. Microsoft also offers Anthropic models inside Copilot. The difference is the enterprise agreement and the settings an administrator controls. A business agreement can set how long prompts are kept, where they are processed, whether they can be used for training and what audit logs are available. A consumer account runs on the provider’s standard terms.

Under the GDPR, that difference matters. If staff send personal data to an AI service, the organization needs a lawful basis, a data processing agreement where the provider acts as processor, a basis for any transfer outside the EEA and security measures appropriate to the risk. A personal ChatGPT account used for work typically provides none of those. Our guide to ChatGPT data security and privacy explains the gap between consumer and business plans.

Lessons for private companies

The Swedish decision is a public-sector move, but the logic applies anywhere staff handle customer, patient or employee data. The same problem shows up in regulated industries, as our piece on why shadow AI is a GLBA Safeguards Rule problem explains for financial institutions. Five steps follow from Sweden’s approach:

1. Find out what is in use. The Transport Administration found about ten services. Most organizations would find more if they looked at network and expense data.
2. Approve tools by agreement, not by brand. Decide which services have contracts that meet your data protection and security needs, and say so in writing.
3. Offer a sanctioned alternative before blocking. Sweden gave staff Copilot first. Blocking without an alternative pushes use onto personal phones.
4. Tighten the approved tool. Review retention, training, connector and sharing settings on whatever you allow.
5. Train people on inputs and outputs. Rules on what not to paste and how to check answers address the risks the auditors actually found.

The limits of a block

A browser block on work computers does not stop someone using a chatbot on a personal device, and it does not catch AI features that arrive inside software the organization already uses. Blocks work best as one layer, alongside policy, training and logging. They also carry a cost: staff who find the approved tool weaker may look for workarounds. Clear guidance about which tasks the approved tool should handle helps reduce that pressure.

Public bodies elsewhere are working through the same questions. Ontario’s updated public-sector privacy impact assessment guide is one example of a regulator asking agencies to assess new technology before deployment. Sweden’s privacy regulator, IMY, has a track record of scrutinizing large platforms, as its consent reprimands show.

Where Captain Compliance fits

Rules on AI use only work if people know them. Captain Compliance’s awareness and training offering helps organizations explain privacy duties to staff, and our assessments and data mapping tools help record which AI services are approved, what data they receive and under what agreement. That record is what a regulator or auditor will ask for. We do not detect or block unapproved AI tools on employee devices, and no training program guarantees that every employee follows the rules. Workplace AI also raises its own employee privacy issues, as the reports about Meta’s exposed employee keystroke data showed.

Frequently asked questions

Has Sweden banned ChatGPT and Claude nationwide?

No. The block applies to staff computers at the Government Offices. Some other agencies, including the Tax Agency and SiS, have their own blocks. Members of the public are not affected.

Why is Copilot allowed?

The Government Offices has an agreement covering Copilot and reviewed and tightened its security settings. ChatGPT and Claude were used without such an agreement.

Can any government staff still use ChatGPT or Claude?

Yes. A limited group can reach OpenAI and Anthropic models through an internal AI platform built by the Government Offices.

What did the Transport Administration audit find?

Widespread, risky use of around ten AI services, with problems in both the sensitivity of information entered and how answers were handled.

Should private companies ban public chatbots?

That depends on the data staff handle and the agreements in place. Many companies take Sweden’s route: approve tools covered by an enterprise agreement, configure them carefully and restrict the rest.

A question of paperwork, not models

Sweden’s government did not decide that one chatbot is smarter or safer than another. It decided that it would only send its data where a contract says what happens next. For most organizations, that is a more practical test than any benchmark, and it raises an uncomfortable question: how many of the AI tools your staff used this week would pass it?