Swedish government offices to ban ChatGPT and Claude

TL;DR: Sweden’s Government Offices will block the browser versions of ChatGPT and Claude on staff work computers from Monday, 26 October 2026, and point employees to Microsoft Copilot instead. The stated reason is information security — staff were using AI services with no contract behind them. It is not a ban on the models themselves: some staff still reach ChatGPT and Claude through a contracted internal platform, which makes this the clearest example yet of a government deciding that the door matters more than the model.

What the Swedish government offices actually blocked

The body in question is the Regeringskansliet — usually translated as the Government Offices. It is not a single ministry but the administrative machine that supports the prime minister and the cabinet, bundling the Prime Minister’s Office with the ministries under one roof. When it changes a rule about work computers, it changes it for the people drafting legislation.

From Monday, 26 October 2026, the browser interfaces for ChatGPT and Claude will no longer be reachable from the service computers (tjänstedatorer) issued to staff. The block was first reported by Sveriges Radio’s Ekot on 23 October at 10:20 UTC, and the offices’ press service confirmed it to broadcaster SVT by email. The scope is narrow and specific, and it is worth being precise about that, because a great deal of the coverage reduced it to “Sweden bans ChatGPT”.

Here is what the reported change looks like route by route:

Three things follow from that table. First, this is a workplace restriction, not legislation — no Swedish law changed, and nothing stops a Swedish citizen from using either product on a personal phone. Second, it targets public web interfaces on managed devices, which is a much smaller surface than “the models”. Third, the exception is not a loophole someone forgot to close. It is the whole point of the exercise.

Why the Swedish government offices blocked ChatGPT and Claude

The reason given is information security, and the specific mechanism is the absence of an agreement.

Mikael Wettercrantz, the Government Offices’ director of digitalisation, declined an interview with Ekot. He did address it internally, writing on the offices’ intranet that using AI services without an agreement can have “consequences for our information security”. That single line is the entire public justification, and it is more revealing than it sounds.

A consumer account on ChatGPT or Claude is governed by consumer terms of service. A procured deployment is governed by a contract, and a contract is where you negotiate the things that actually matter to a public body:

• Retention. How long are prompts and outputs stored, and can the customer set that to zero?
• Processing location. Which country’s servers handle the request, and does the data leave the EU?
• Training. Are inputs used to improve the model, and can that be switched off contractually rather than by a checkbox?
• Sub-processors. Who else touches the data, and does the customer get notice before that list changes?
• Administrative control. SSO, device-level policy, the ability to disable features, and audit logs tying a session to a named officer.
• Breach notification and liability. Who tells the regulator, within what window, and who carries the cost.

None of that exists when an officer opens a browser tab on a personal account. The prompt goes to the provider’s servers under standard terms, and the agency has no visibility.

Ekot’s reporting also points to a concrete trigger. Earlier in 2026, an internal audit at Trafikverket, the Swedish Transport Administration, found widespread and risky use of around ten different AI services. The audit’s own words, as quoted by Ekot: “We have taken note of several examples of problematic use of open AI services. The examples concern both the sensitivity of the information fed in and how the AI-generated answers were handled.”

That second clause is the part people skip. It is not only about what goes in. It is about what comes back out and what a human being does with it — an AI-drafted paragraph pasted into an official document without verification is its own class of risk, and no data processing agreement fixes it.

Copilot stays, and the reason has nothing to do with model quality

If the concern were the underlying technology, approving Microsoft Copilot would be strange. Copilot is a general-purpose assistant built on frontier models, and Microsoft has been offering Anthropic’s Claude models inside it since early 2026.

The distinction the Government Offices draws is procedural. Copilot was reviewed, and then its security settings were raised as a result of that review — something the offices did not do for ChatGPT or Claude. When the press service was asked what the hardening involved, it declined to give details, citing security.

That gap between the two treatments maps onto a distinction most organisations never write down: a consumer subscription and an enterprise deployment of the same product are not the same product.

There is a nice irony buried in the Swedish decision. In the EU, Anthropic’s models are off by default inside Copilot and have to be deliberately enabled by an administrator. The Government Offices has not said whether it enabled them. So it is entirely possible that Claude the website is blocked on a Monday morning while Claude the model remains quietly available inside the one tool the agency just approved — and that this is not a contradiction, because the thing being governed was never the model.

Intric: keep the models, change the door

The exception route is the most instructive part of the story, and it is also the part reported with the least precision.

Sveriges Radio’s press service says part of the staff will keep access to a special platform for AI assistants, which the offices developed themselves, and through which they can reach ChatGPT and Claude. Other reporting identifies the platform’s supplier: Intric, a Stockholm-based AI platform company, which announced on 8 July 2026 that it had won the Government Offices’ procurement for a platform where staff build and run AI assistants on the agency’s own data. Intric describes the platform as running on Swedish infrastructure “without dependence on foreign cloud services”, and the agreement reportedly covers both a cloud service and an on-premises installation.

Both things can be true at once — an agency can run its own build on top of a procured platform — but neither the offices nor Intric has publicly said which deployment is in production, how many staff are covered, or where requests to the ChatGPT and Claude models are actually processed. Treat any confident claim about the internals as unverified.

What is clear is the shape of the decision:

1. Close the public websites, where nobody signed anything.
2. Keep the models, but only through a platform the agency contracted, configured and can audit.
3. Tell staff which door is open, so the rule is followable rather than merely prohibitive.

That is a governance pattern, not a technology ban, and it is one that will travel. Every organisation running AI on personal accounts with nothing signed is one policy review away from a block landing on a Monday morning.

Other Swedish agencies are already there

The Government Offices is late to its own party. Ekot’s reporting found blocks already in place at two other agencies and a third preparing a decision:

• Skatteverket, the Swedish Tax Agency, already blocks consumer AI tools. Ekot did not say when the block started or exactly which services it covers, so treat that entry as the least specific.
• Statens institutionsstyrelse (SiS), the National Board of Institutional Care, which runs youth homes and addiction treatment facilities, also has a block in place. Of all the agencies listed, SiS has the strongest case: its staff handle the personal data of minors in state care, which is about as sensitive as Swedish public administration gets.
• Kronofogden, the Swedish Enforcement Authority, is preparing its own position.
• Trafikverket has not blocked anything publicly, but its internal audit is the evidentiary backbone of the whole story.

Outside Sweden, the European Parliament has taken a comparable step, restricting AI features on staff devices over data-security concerns. The direction of travel across European public administration is consistent: not “no AI”, but “no AI without a signature”.

The apparent contradiction: restricting tools while chasing AI leadership

It would be easy to read this as Sweden going cold on AI. The record says otherwise. Sweden published its first comprehensive national AI strategy in February 2026, with an explicit ambition to rank among the world’s leading AI nations, and the country’s Agency for Digital Government (Digg) has issued joint guidance with the privacy regulator IMY on generative AI in public administration — covering management responsibility, procurement, information security, confidentiality and human control.

That is not a contradiction; it is a sequencing choice. A government that wants 300 agencies to adopt AI has to make adoption legible to auditors. If it lets every officer pick their own chatbot, it cannot answer the first question any parliamentary committee will ask: what citizen data went where. Standardising the door is the precondition for scaling the use.

The cost is real, though, and worth naming. Centralising the route also centralises the failure: if the approved platform is bad, slow, or badly suited to a task, staff do not simply stop working — they route around it, and shadow AI is harder to find than a blocked website. The Government Offices has not published usage data, an exception process, or any measure of whether Copilot actually covers the tasks the block removed. Until it does, “the block works” is a claim about a firewall, not about information handling.

What this means for Malaysian developers and public-sector teams

For a Malaysian reader, nothing about this changes what you can do on your own laptop tomorrow. Its value is as a preview of the procurement questions that are coming, because Malaysia has already built most of the policy scaffolding that Sweden is now enforcing against.

Malaysia’s framework already exists. The National Guidelines on AI Governance and Ethics (AIGE) set out seven principles — fairness; reliability, safety and human control; privacy and security; inclusiveness; transparency; accountability; and human-centricity. The Public Sector AI Adaptation Guidelines, launched on 27 February 2025 by the National Digital Department (JDN) with MDEC, translate that into agency-level practice across six chapters and three appendices, including risk management and adaptation methods. The National AI Office (NAIO) was established in December 2024, the AI Technology Action Plan 2026–2030 carries an explicit “AI-Competent Civil Service” pillar, and the PDPA amendments passed in 2024 added mandatory breach notification, data protection officer requirements and tighter rules on cross-border transfers.

What Malaysia has not published, as far as I can find, is a Swedish-style block. There is no public record of a ministry cutting off ChatGPT or Claude on staff machines. That is a difference in enforcement, not in direction, and it is the gap a vendor or a systems integrator should plan for.

Three practical implications:

1. If you sell into the public sector, expect the contract to be the product. The first questions in a government tender will not be about benchmark scores. They will be: where does inference run, what is the retention period, can training on our inputs be switched off in writing, who are the sub-processors, and can we get an audit log per officer. Have answers ready before you are asked.
2. API access and website access are different products with different fates. Sweden blocked websites. Contracted and API routes survived, including routes that reach the identical models. If you are building a product on OpenAI or Anthropic for enterprise customers, the commercial terms — zero data retention, regional processing, DPA — are not paperwork. They are the part of your offering that survives a policy review.
3. Malaysian data residency is now a selling point. AWS, Google Cloud and Microsoft Azure all operate Malaysian regions. A deployment that keeps inference in-country is an easier conversation with a Malaysian agency than one that does not, and it is the local equivalent of Intric’s “Swedish infrastructure” pitch.

A five-step audit any team can run this week

The Swedish story is really an audit finding that got published. Any organisation can run the same check on itself, and the output fits on one page.

1. Inventory the routes, not the tools. List every way work data can reach a model: browser tabs, desktop apps, IDE plugins, browser extensions, meeting transcribers, and the “AI” button inside products you already license. Trafikverket found about ten; most teams underestimate by half.
2. Classify the data that flows through each route. Public, internal, personal data, and regulated material are four different answers. A route that is fine for a published consultation paper is not fine for case files.
3. Check whether a contract exists. Not a free tier, not a trial, not a personal Plus subscription expensed on a company card. An agreement with retention, location, training and breach terms.
4. Route everything through one controlled door. One approved platform with SSO, logging, and a configuration you can prove. Publish the allowlist so staff know what “approved” means.
5. Log and review. Without usage data you cannot tell a working policy from a decorative one.

A machine-readable allowlist makes step four enforceable rather than aspirational:

# ai_policy.yaml — one approved door per data class
version: 2026-10
default: deny

routes:
  public_web_chat:
    tools: [chatgpt.com, claude.ai, gemini.google.com]
    allowed_data: [public]
    contract: none
    requires_review: true

  approved_assistant:
    tools: [copilot_enterprise, intric_internal_assistant]
    allowed_data: [public, internal]
    contract: dpa-signed
    retention_days: 0
    training_on_inputs: false
    processing_region: my-central
    audit_log: required

  regulated_casework:
    tools: [internal_assistant_on_prem]
    allowed_data: [public, internal, personal, regulated]
    contract: dpa-signed
    human_review: mandatory
    export_blocked: true

The Swedish version of this file exists; it just is not public. What is public is the decision it produced: deny the open routes, allow the signed one, and tell staff which is which.

FAQ

Is Sweden banning ChatGPT and Claude?
No. The Government Offices is blocking the public websites on its own staff work computers. It is not a law, it does not apply to Swedish citizens or private companies, and the models themselves remain reachable through an approved internal platform for some staff.

When does the block take effect?
Monday, 26 October 2026, according to the Government Offices’ email to SVT. Ekot reported the decision on 23 October 2026.

Can Swedish government employees still use Claude?
Some can. The press service says part of the staff retains access to a dedicated AI-assistant platform through which they can reach both ChatGPT and Claude. It has not said how many people that covers or what restrictions apply inside it.

Why is Microsoft Copilot allowed if ChatGPT and Claude are not?
Because the offices reviewed Copilot and then raised its security settings, something it did not do for the other two. The distinguishing factor is the contract and the configuration, not the quality of the model — and Microsoft has offered Anthropic’s Claude models inside Copilot since early 2026.

Does this affect Malaysian users or businesses?
Not directly. It is a workplace policy in one country. What travels is the template: if your team uses AI on personal accounts with no agreement behind it, an audit can produce the same outcome anywhere, and Malaysian public-sector procurement increasingly asks the same questions Sweden just answered.

The bottom line

Sweden’s Government Offices did not decide that ChatGPT and Claude are dangerous. It decided that an unaccountable route into a frontier model is unacceptable for civil servants handling other people’s data — and then built an accountable one, kept the models, and told staff where the door is.

That is a more mature position than either a ban or a free-for-all, and it is the one most large organisations will eventually land on. The lesson for anyone building or buying AI tools is uncomfortable but simple: the model is not the thing being governed. The contract, the configuration and the audit trail are. Teams that can produce those three documents keep their tools. Teams that cannot get a block on a Monday morning.

More from ZFRBuild

• Malaysia’s busiest stretch: what to book before 5 December
• Alternatives to Animal Testing Are Finally Taking Off
• Dryers buying guide: what actually matters in Malaysia