From 26 October 2026, Sweden's Government Offices (Regeringskansliet) blocks browser access to ChatGPT and Claude on staff computers. Microsoft Copilot is the approved alternative, and the underlying models remain reachable through Intric, a platform used across the ministries, according to CEOWORLD's report citing SVT. The decision is narrower than the headlines suggest, and better for it.
The strongest case for the block
The security argument deserves a fair hearing. The Government Offices' director of digitalisation, Mikael Wettercrantz, told staff on the intranet that using AI services without a formal agreement creates information-security risk. He also said employees had entered sensitive information into open AI services and mishandled AI-generated answers. A ministry that supports a prime minister and cabinet handles material whose exposure is costly: negotiating positions, security assessments, personal data. Consumer terms of service let a vendor change the rules unilaterally and give the customer no audit rights. Blocking at the browser is a blunt tool, but it is the only control that works when staff are already pasting documents into chatbots.
The pattern is not unique to Stockholm. According to the same report, the Tax Agency and the National Board of Institutional Care have introduced similar blocks, and the Enforcement Authority is considering one. An internal audit at the Transport Administration reportedly found about 10 AI services in "widespread and risky" use. That is shadow IT at scale, and it is the real problem.
What the block actually distinguishes
The notable feature is what Sweden did not do. It did not declare ChatGPT or Claude unfit for government. The same models stay available through Intric, and the Government Offices approved Copilot after reviewing it and tightening its security settings before a rollout to all employees. The reported distinction is contractual: an enterprise agreement with controls the ministry negotiated versus a consumer login nobody negotiated. As the CEOWORLD editor put it, "Sweden has not rejected any AI model. It has rejected ungoverned access."
That framing matters because it separates two policies that are often conflated. A capability ban says a technology is too dangerous for public servants. A procurement control says a technology must arrive through a channel with defined data handling, logging and liability. The first slows public-sector productivity and pushes use onto personal phones, where there is no oversight at all. The second preserves the benefit and shrinks the risk.
Where the policy is vulnerable
A pro-innovation reading should still note the weak points.
- Vendor lock-in risk. Approving one vendor by default concentrates dependence on a single U.S. company. The reporting says the approval rests on the existing Microsoft enterprise agreement rather than any jurisdictional difference, since Microsoft is also American. If so, the security gain is contractual, not sovereign, and should be described that way.
- Unanswered transparency questions. The press office did not say how many employees can use Intric or what restrictions apply. Without published criteria, other agencies cannot copy the approach, and vendors cannot compete for approval on the merits.
- Browser blocks are leaky. They stop casual use on managed machines, not copy-and-paste to a personal device. Training and a usable approved tool do more than a firewall rule.
The wider European context
Sweden is not acting in a vacuum. In February 2026 the European Parliament disabled built-in AI features, such as writing assistants and summarisers, on lawmakers' and staff devices because its IT team could not guarantee the security of data sent to external providers, as recorded in the OECD AI Incidents and Hazards Monitor. The OECD classes this as a hazard, not an incident: no harm was reported. Public bodies are acting on plausible exposure, not demonstrated leaks.
The regulatory backdrop is also moving. Per the Commission's AI Act page, the Act became applicable on 2 August 2026, with the AI Office and national authorities taking on supervision and transparency rules taking effect, while high-risk obligations phase in from December 2027. Separately, the European Data Protection Board adopted Opinion 28/2024 on 18 December 2024, addressing data protection aspects of processing personal data in AI models. Neither instrument tells a ministry which chatbot to allow. That decision stays with the deploying organisation, which is why internal procurement discipline matters more than any single statute.
What proportionate looks like
Sweden's approach points to a workable template for other governments.
- Publish the criteria. State what an approved tool must provide: data residency terms, no training on inputs, logging, exit rights. Then any vendor, European or American, can qualify.
- Provide the approved route first. Block consumer access only once a sanctioned alternative is live, as Sweden did. Prohibition without substitution drives shadow use.
- Keep model choice open. Routing several models through a platform like Intric lowers lock-in better than a single-vendor mandate.
- Measure, don't assume. The Transport Administration's audit counted the services actually in use. Every ministry should do the same before drafting rules.
The risk is that other capitals read Stockholm's decision as permission for blanket bans. The evidence supports the opposite conclusion. The fault was the absence of a contract, and the fix is to supply one. Governments that treat AI as a procurement discipline will capture productivity gains and protect sensitive data. Governments that treat it as a prohibition will get neither.