Sweden's Government Offices to block ChatGPT and Claude on work computers, passes Microsoft Copilot as the approved option
Sweden's Government Offices will block ChatGPT and Claude on staff work computers from October 26, per Ekot, steering staff to a reviewed Microsoft Copilot.
Sweden’s Government Offices — Regeringskansliet, the joint administrative body serving the country’s ministries — will block OpenAI’s ChatGPT and Anthropic’s Claude on staff work computers from Monday, October 26. Sveriges Radio Ekot, which obtained internal information from the Government Offices, reported on October 23 that the two services “will no longer be reachable from staff work computers” starting Monday, with employees steered toward Microsoft Copilot, the office’s approved assistant. The Tax Agency (Skatteverket) and the National Board of Institutional Care (Sis) already run similar blocks. As of publication, the Government Offices had issued no press release about the ban; the entire story rests on Ekot’s sourcing.
Key points
- The block covers website access to ChatGPT and Claude on Government Offices work computers, effective Monday, October 26. Ekot broke the story on October 23; the TT wire and SVT both followed the same day, each attributing everything to Ekot.
- Mikael Wettercrantz, the office’s director of digital affairs, declined an interview and left a comment on the intranet instead: AI services used without agreements can have “consequences for our information security.” Ekot relayed the quote; the press office refused to elaborate, citing security.
- Microsoft Copilot was not waved through alongside the ban. According to the internal information Ekot saw, the office ran a dedicated review of Copilot and raised its security configuration. Some staff will keep access to ChatGPT and Claude through an AI platform the Government Offices built itself.
One source, many bylines
The first thing to keep straight: this is a reported leak, not an official announcement. Ekot’s two hardest facts — the effective date and the intranet quote — come from internal information its reporters obtained. Wettercrantz declined to be interviewed, the press office declined to give details, and SVT’s own version simply ends with “SVT Nyheter is seeking comment from the Government Offices.” TT’s wire copy as carried by Ny Teknik, and SVT’s report, both anchor every claim to Ekot without adding independent reporting of their own; Börsvärlden runs Finwire’s rewrite of the Ekot story, with no added sourcing. The underlying source count is still one.
The word “approved” needs the same care. SVT states that Copilot has been cleared by the Government Offices “after a review of the service raised its security” — a formulation that again traces back to Ekot’s internal information, not to any statement from Microsoft or the office itself. What the review tested, which settings changed, and who signed off: none of that is externally checkable. By the standard this site applies to news, every fact in the chain carries the tag “according to internal information seen by Ekot.”
The facts
Read together, Ekot, TT and SVT yield this set of checkable claims:
| Item | Detail | Source |
|---|---|---|
| Block takes effect | 2026-10-26 (Monday): work computers can no longer reach ChatGPT or Claude | Ekot, via TT |
| Quote | Wettercrantz: uncontracted AI use can have “consequences for our information security” | Ekot (intranet) |
| Replacement | Microsoft Copilot, listed as approved after a dedicated security review | Ekot internal info, via SVT |
| Exception | Some staff reach ChatGPT and Claude through the office’s own AI platform | Ekot |
| Other agencies | Skatteverket and Sis already block; Kronofogden is preparing a decision | Ekot survey |
| Transport audit | Trafikverket’s internal audit found problematic use of about ten public AI services | Ekot only |
| DeepSeek | Already restricted before this ban | Ekot only |
Two details come from Ekot alone, and deserve their provenance stated. First, the Trafikverket (Swedish Transport Administration) audit: earlier this year it found roughly ten public AI services in problematic use, covering both the sensitivity of information fed into them and how the AI-generated answers were handled. Second, the earlier DeepSeek restriction inside government — the report gives it one sentence, with no date or scope. Neither detail made it into TT’s or SVT’s coverage.
How Swedish agencies compare
Ekot’s survey sketches an incomplete map of AI blocks across Swedish government: the Government Offices joins on October 26; the Tax Agency and Sis, which oversees institutional care, acted earlier; the Enforcement Authority (Kronofogden) is preparing its own decision. There is no single decree — each agency inventoried its own tools and moved on its own schedule, and the offices’ own net only closes on Monday.
The stagger itself is informative. No agency waited for central direction; security officers moved first. The justification the Government Offices offered — uncontracted services risk leaking information — turns on the contract, not the model: with no agreement governing where the data goes, the risk is treated as established. How the rest of Europe draws this line is beyond what this entry could verify from primary material.
Our take
For teams selling to European public sector, the signal matters more than the story: procurement is shifting from “does it work” to “who reviewed it, and on what basis.” The Government Offices’ choice is blunt — not the strongest model, the reviewed vendor. Copilot wins on compliance status, not capability ranking, and that is the second half of a story we have tracked: Microsoft’s rework of Copilot into the “OS for work” is the same strategy seen from the product side (Microsoft rethinks Copilot). ChatGPT and Claude lost here on enterprise agreements and data-residency defaults, not on benchmarks.
For the employees living under the block, the picture is looser than the headline. It covers work computers only, not phones or private devices, and the office itself is keeping an exception channel open for some staff — an implicit admission that “no AI at all” was never realistic. Open-source governance is tracing the same arc: Zig banned AI-generated contributions while conceding private use is unenforceable (Zig bans AI contributions). The line is moving from “whether you may use it” to “what counts as sanctioned use.”
The caveats are just as clear. Ekot remains the sole source; the office will not even discuss details, so “Copilot’s security was raised” is an internal claim for now. How the block behaves in practice, and whether any agency reverses course, waits on the next round of reporting. The takeaway for vendors, though, needs no follow-up: have your security-review paperwork ready to hand, because your next public-sector procurement list is likely to carry the same gate.