Sweden Government Blocks ChatGPT and Claude on Work PCs, Approves Microsoft Copilot

Some headlines will call this "Sweden bans ChatGPT." It isn't that. It's an access block on government-issued machines, with an approved alternative and a controlled exception. The decision is still a clear example of where enterprise AI governance is heading, and Windows admins and Microsoft 365 tenant owners should pay attention.

What Regeringskansliet actually decided​

The facts reported so far are narrow:

• Scope: Regeringskansliet staff will no longer be able to reach the AI tools from their service computers. The reports cover work devices only. They say nothing about personal phones or home PCs.
• Rationale: Regeringskansliet wants staff to stick to the AI services the organisation has approved, according to Ekot.
• The approved option: Copilot is described as approved by Regeringskansliet, which raised its security after reviewing the service.
• Secrecy about the Copilot changes: the tightened Copilot settings appear in internal information Ekot has seen, but the press office won't go into detail because it concerns security.
• An exception: According to the press office, some staff keep access to ChatGPT and Claude through an AI-assistant platform Regeringskansliet built itself.
• Earlier action: Regeringskansliet had already restricted use of DeepSeek.

The digitalisation director, Mikael Wettercrantz, declined an interview with Ekot. On the Government Offices' intranet, he wrote that using AI services without an agreement could have "consequences for our information security." Feber noted that Regeringskansliet doesn't want to discuss with the media exactly why it is blocking the two services.

The key phrase is "without an agreement." The official reason isn't that ChatGPT or Claude are dangerous technology. It's that staff were using them outside any contract the organisation had signed. That fits with the exception: the same models stay available to some staff through an in-house platform the organisation controls.

Section summary: Direct browser access to ChatGPT and Claude ends on work PCs. Copilot is the approved default, a controlled internal route to both blocked models remains for some staff, and the specific Copilot hardening is not being disclosed.

Not an isolated decision​

Other Swedish agencies have done the same. Ekot reports that similar blocks are already in place at Skatteverket, the Swedish Tax Agency, and at the National Board of Institutional Care (SiS), and the same may happen at Kronofogden, the Swedish Enforcement Authority. To be precise, Kronofogden had not decided anything when the story ran. The question was still being prepared.

The most concrete evidence comes from an internal audit at Trafikverket, the Swedish Transport Administration. Ekot reports that this year's audit found staff using about ten different open AI services. The auditors wrote that they had seen several examples of problematic use, involving both how sensitive the entered information was and how employees handled the AI-generated answers.

The audit raises two separate problems:

1. Input risk: sensitive information going into services the agency has no contract with.
2. Output risk: staff trusting or reusing AI answers carelessly.

Blocking a website mostly deals with the first problem. The second one, people treating a chatbot's confident paragraph as settled fact, will follow employees to whichever tool is approved, Copilot included.

The reporting does not say a breach occurred, what kind of data was entered, or which ten services were involved. Treat the audit as evidence of risky habits, not a confirmed leak.

Section summary: At least three Swedish public bodies have blocked or are considering blocking open AI services. The one detailed finding describes sensitive inputs and careless handling of outputs, not a confirmed data loss.

The Swedish legal backdrop​

Swedish law adds a complication that most corporate IT departments don't face: the offentlighetsprincipen, or principle of public access to official records.

On September 3, 2026, the Ministry of Justice announced that a special investigator would analyse how that principle applies when authorities use AI and external cloud services. The ministry wrote that AI tools handle, store and generate information in ways that can mean information passes outside an authority. That creates uncertainty about when such material counts as an official public document. The investigator, Lisa Englund Krafft, will examine:

• how authorities use AI, including tools in external cloud services
• whether information generated or stored there counts as an official document
• whether current rules hold back adoption

The report is due by November 15, 2027.

The government hasn't linked this inquiry to Monday's block, and nothing says the block is legally required. Still, if nobody can say whether a prompt typed into an uncontracted chatbot is a public record, an agency without a contract can't easily answer a records request for it. Keeping AI use inside contracted services is one way to reduce that uncertainty.

Section summary: Sweden is formally studying whether AI-generated and cloud-stored material counts as public records. The findings are not due until late 2027, which helps explain why agencies are acting cautiously now.

Sweden isn't anti-AI​

This isn't a government turning against AI. In an October 2025 summary of reports from 25 agencies, the Ministry of Finance said agencies were actively exploring AI but many were still at an early stage. Examples included:

• The Swedish Post and Telecom Authority (PTS) ran Copilot and ChatGPT pilots for a number of staff under strict information-handling guidelines.
• Kammarkollegiet tested ChatGPT and Copilot and began rolling out digital assistants to all employees.
• The National Institute of Economic Research described tools such as Copilot as cost-effective. It judged the information-security risk very low because of its technical environment and data routines.

The lesson is that the same chatbot can be fine in one organisation and off-limits in another. What decides it is contracts, configuration and training, not the product alone.

Why Microsoft wins by default, and what that does and doesn't prove​

Copilot comes out of this as the approved assistant, which is good news for Redmond. Sejfer noted that Copilot stands apart here because Regeringskansliet approved it after its review and took steps to raise its security.

That shouldn't be read as an independent ruling that Copilot is technically safer than ChatGPT or Claude. The reports contain no comparative security assessment. Regeringskansliet hasn't said which Copilot edition it uses, how its tenant is configured, or what contract terms apply. The most likely explanation, based on general industry experience, is simpler: organisations already running Microsoft 365 usually have a contract, identity integration and compliance tooling in place, so the bundled assistant is the easiest tool to approve.

Two caveats:

• Approved isn't the same as authorised for everything. Approved status doesn't mean any type of classified or personal data may go into Copilot. Data-classification rules still apply.
• Hardening after a review implies something needed hardening. Regeringskansliet tightened Copilot's settings only after reviewing it. Out-of-the-box defaults may not have met its requirements, which is worth noting for any tenant that switched Copilot on and moved on.

What enterprise IT admins can take from this​

Regeringskansliet hasn't said how it is enforcing the block. It could be network filtering, browser policy, device management or identity controls. The checklist below is based on general industry practice, not on its implementation:

1. Inventory shadow AI first. Trafikverket found about ten services in use. Your proxy, DNS or cloud-app discovery logs will probably show a similar spread. Know what's in use before you block anything.
2. Publish an approved list with reasons. Staff accept a block more easily when they know what to use instead and why.
3. Define which data may go into which tool. A blocklist does nothing about someone pasting a confidential memo into the approved assistant.
4. Review your Copilot configuration rather than trusting defaults. Check permissions and oversharing in SharePoint and OneDrive, sensitivity labels, and data loss prevention coverage. Copilot surfaces whatever a user can already access.
5. Offer a controlled route for legitimate needs. Regeringskansliet's in-house platform is effectively a sanctioned gateway to the blocked models. A block with no alternative tends to push people to their personal phones.
6. Train people on handling outputs, not just inputs. That was half of Trafikverket's finding, and no firewall rule fixes it.

Section summary: A blocklist is the easy part. Discovery, data rules, a properly configured approved tool, a sanctioned exception path and training on outputs are what make an AI policy work.

The bottom line​

Regeringskansliet's move is less a verdict on OpenAI or Anthropic than a statement about control: AI is allowed, but only through services the organisation has a contract with and has configured. Copilot benefits because Microsoft already has the contract and the tenant. Keeping a controlled internal route to ChatGPT and Claude suggests officials still see value in those models when they control the environment.

Open questions remain. Will the block push some staff onto personal devices? Will other agencies follow Skatteverket and SiS? And what will the 2027 public-access inquiry decide about the prompts and answers already generated in government systems?

For admins elsewhere, the practical point is simple: the first step in an AI policy is knowing which tools people are actually using.