Sweden's Government Offices — Regeringskansliet, the chancery that supports the prime minister and the ministries — begin blocking ChatGPT and Claude on staff work computers on Monday, 26 October. Sveriges Radio's Ekot programme reported the decision on 23 October, and SVT News confirmed the details.
The block applies to the office's own service machines. It is not a law, it is not a national ban, and it does not forbid generative AI as a category: staff who need an assistant are directed to Microsoft Copilot, which the chancery says it has approved after a review of the service and a tightening of its security settings.
A pattern, not a one-off
Regeringskansliet is not first. Blocks have already been imposed at Skatteverket, the Swedish Tax Agency, and at Statens institutionsstyrelse (SiS), the agency that runs the country's youth homes and secure institutions. Kronofogden, the Swedish Enforcement Authority, is reported to be considering the same step.
That sequencing is the news value here. Sweden's public administration is not running a single dramatic reversal of AI adoption; it is grinding through a procurement-and-classification exercise agency by agency, and the two best-known American chat assistants are failing it while an incumbent enterprise product passes. Microsoft Copilot's advantage in that comparison is not capability — it is that a government body can point to a data-processing agreement, an assessed security posture and a named counterparty.
What the staff were told
The internal justification, as reported, is information security. Ekot's reporting describes the use of the two services as problematic from an information-security standpoint; Swedish outlet Nyheter24 reported that some staff may still reach both tools through a platform the chancery built itself.
An SVT request for comment from Regeringskansliet was still pending when it published, and neither OpenAI nor Anthropic had issued a statement on the block at the time of writing. Treat the exact internal reasoning as reported rather than officially confirmed.
Why procurement keeps beating product
For anyone selling AI into the public sector, the Swedish case is a tidy illustration of where deals are actually decided. ChatGPT and Claude are blocked on the basis of contractual and jurisdictional questions — where the data is processed, under which agreement, with whom liability sits — not because a civil servant judged the answers worse. Copilot wins those rooms because it arrives inside an agreement the organisation already signed.
That also sets the shape of the fix. A vendor that wants Regeringskansliet's users back does not need a better model; it needs a Swedish or EU data-processing arrangement, a security review it can submit to, and a name on the contract. The same logic is already visible across European administrations, where the shortlist is usually drawn around data residency rather than benchmark scores.
The practical consequence for the ~2,000-odd people working in the chancery is unglamorous: the assistant they were using disappears from their taskbar on Monday morning, and a different one is already installed in its place.
