Sweden Bans ChatGPT and Claude, Blesses Copilot Instead
Sweden Bans ChatGPT and Claude, Blesses Copilot Instead
By Daniel Rao · Stockholm — October 23, 2026
Last August, Sweden's prime minister cheerfully told a business newspaper that he consults ChatGPT for second opinions on political matters. Come Monday morning, the government offices he leads will firewall ChatGPT — and Anthropic's Claude — off every staff work computer in Stockholm. Sweden is not turning its back on AI. It is drawing a hard line around which AI its civil servants may talk to, and that line runs straight through Redmond.
According to reporting by Sveriges Radio's Ekot on Friday, the Government Offices of Sweden — the Regeringskansliet, the administrative engine room behind the country's ministries — will implement a hard network block on the two chatbots starting October 26. Staff who type chat.openai.com or claude.ai into their office browsers will get nothing back. The only chatbot left standing will be Microsoft's Copilot, which the offices have formally approved as the sole sanctioned AI tool.
The Monday switch-off
Let us be precise about what is — and is not — happening. This is not a national ban. No Swede will lose access to ChatGPT on their phone or home laptop. What the Government Offices are doing is an employer move: a network-level block applied to staff service computers, the machines civil servants use for their daily work.
From Monday, direct web access to OpenAI's ChatGPT and Anthropic's Claude goes dark on those machines. There are carve-outs. A slice of personnel will keep reaching the same underlying models through an internally developed platform built to interface with ChatGPT and Claude in a controlled way — queries fenced in, data flows the offices can see. Everyone else gets Copilot or nothing.
The driving force is not some newly discovered vulnerability. Internal documents seen by Ekot show the offices want staff locked onto tools the government actually has agreements with — enterprise contracts with defined data handling, audit trails, and somebody to sue if it all goes wrong. The consumer chatbots, whatever their merits, offer none of that on paper.
A terse intranet note and a declined interview
The most revealing detail of Ekot's reporting is not the block itself. It is how the block was communicated.
Mikael Wettercrantz, the Government Offices' digitalization director, declined Ekot's request for an interview. Then he wrote a note on the agency's intranet warning that AI services used without an agreement can have "consequences for our information security." That is the entire public-facing rationale: a refused interview and a terse internal post, delivered to staff rather than to the press.
We checked the Government Offices' public press pages on Friday — there is nothing about the block posted there. The news broke on the intranet first, and the government has so far said nothing to the outside world. The press office, asked about the simultaneous security upgrade to Copilot, declined to describe the technical changes, citing internal security protocols. Which is fair enough, as far as it goes — but it leaves everyone guessing about what exactly was hardened, and whether it was hardened enough.
The firewall rule is the message.
From second opinions to firewalls: the Kristersson irony
And then there is the prime minister.
In August 2025, Ulf Kristersson told the business daily Dagens industri that he uses AI tools — ChatGPT and the French service LeChat among them — "quite often," adding: "I use it myself quite often. If for nothing else than for a second opinion. What have others done? And should we think the complete opposite?" His colleagues, he said, use the tools in their daily work too.
It caused a small political storm. Tech experts fretted about politicians outsourcing judgment to machines; the newspaper Aftonbladet accused him of falling for what it called the oligarchs' AI psychosis. Virginia Dignum, a professor of responsible AI at Umeå University, told Dagens Nyheter the habit was "a slippery slope," warning: "The more he relies on AI for simple things, the bigger the risk of an overconfidence in the system." Her kicker line — "We didn't vote for ChatGPT" — made the rounds.
His spokesperson later insisted he never fed the tools anything sensitive: "Naturally it is not security sensitive information that ends up there. It is used more as a ballpark." The subtext of Monday's block is that the government no longer trusts its staff to draw that line on their own — or that it finally tallied what "a ballpark" means when thousands of employees do it daily.
Not Italy 2023 — this is an employer's block, not a nation's
It is tempting to file this alongside Italy's 2023 ChatGPT episode, and the comparison is instructive precisely because it shows how different this is.
In March 2023, Italy's data protection authority, the Garante, ordered a temporary nationwide block of ChatGPT — the first Western regulator to do so — over the alleged lack of legal basis for harvesting personal data to train the model, and over missing age-gating for minors. OpenAI took the service offline for Italian users, made concessions on transparency, opt-outs, and age verification, and the ban was lifted at the end of April. A regulator banned a product for an entire country; the product changed; the ban died.
Sweden is doing nothing of the sort. No regulator is involved. No citizen loses anything. The Regeringskansliet is acting as an employer managing its own network — closer to a bank telling traders they cannot install Telegram on work laptops than to a government banning an app. The legal basis is internal IT policy, not data-protection law, and there is no OpenAI concession that could "lift" it, because the block is not addressed to OpenAI at all.
That distinction matters, and it is worth stating plainly: Stockholm is not banning AI chatbots. It is buying one AI chatbot and uninstalling the others. The question is whether that is security strategy or procurement theater.
Sovereign AI ambitions, rented American plumbing
Here is where the story gets genuinely strange. Sweden is simultaneously one of Europe's most ambitious AI sovereigntists and, as of Monday, a government that runs its AI on an American product.
In February 2026, Stockholm adopted its first national AI strategy, with Kristersson declaring Sweden should rank among the world's top ten AI nations and become the world leader in using AI inside public administration. The strategy commits serious money — SEK 479 million in AI and data reforms this year alone — and its crown jewel is a homegrown Swedish-language model, a "sovereign intelligence" initiative meant to guarantee digital independence and reflect Swedish values in the training data. A National AI Workshop, a shared infrastructure for public-sector AI, is supposed to be fully built out by 2030.
Ten days ago, France's Scaleway announced it would build Europe's first commercial AMD-powered AI supercomputer in Sweden, running AMD Instinct MI355X chips — a bet on Sweden's low-carbon electricity and its AI sector, as we covered at the time. "Scaleway is expanding the choice of AI infrastructure available to customers across Europe, through AMD Instinct GPUs and our open ROCm software ecosystem," Stephanie Dismore, AMD's senior vice president for EMEA, told Reuters. Tele2's Stefan Trampus said it "sends a clear signal" on Sweden's attractiveness for AI infrastructure.
So the pattern is unmistakable: Sweden wants its own models, its own compute, its own stack. And yet, on the most basic question — what chatbot may a civil servant open on a Tuesday — the answer is Microsoft's. There is no Swedish model ready for the ministries. There is no sovereign assistant. There is Copilot, an American product, wrapped in a reviewed enterprise agreement and an upgraded security configuration the press office will not describe.
This is the dirty secret of the global sovereign-AI wave: sovereignty is a decade-long construction project, but Tuesday comes every week. Governments ban the chatbots they cannot control while renting the chatbot they have a contract for, and both moves are made in the name of the same word — security.
The Copilot paradox
Strip away the bureaucracy and the arrangement is this: Sweden's government does not trust ChatGPT or Claude with its data, so it will trust Copilot with its data instead. Copilot is, of course, built on OpenAI's models — the very models whose consumer versions are being blocked. The difference is contractual, not technical: an enterprise agreement, data-residency commitments, audit controls, and a procurement relationship with a vendor the Swedish state has done business with for decades.
That is not nothing. Enterprise AI contracts genuinely change the data-handling picture — inputs walled off from training, retention limits, regional processing. But it does mean the block's real message is narrower than the headlines suggest. It is not "these models are unsafe." It is "these models are unsafe without a contract." Which raises the obvious follow-up: if a future enterprise agreement with Anthropic materialized, would Claude walk back through the firewall? Almost certainly.
The same logic is spreading. The Swedish Tax Agency and the National Board of Institutional Care have already imposed their own AI-service blocks, and the Enforcement Authority is weighing one. The Government Offices had previously restricted DeepSeek, the Chinese chatbot, on its machines — a move that sailed through with far less fuss, because geopolitics makes an easy villain. Blocking the American chatbots while mandating the American enterprise assistant is the harder, more honest version of the same calculation.
Across the Atlantic, the pattern rhymes: this week the US Department of Homeland Security ordered civilian federal agencies to purge DeepSeek from their systems. Governments everywhere are converging on the same posture — use AI, but only the AI with paperwork.
Will the block actually work?
Probably not the way Stockholm hopes — and the government seems to know it, which is why the exceptions exist.
Network blocks on office machines stop the casual, careless use: the paste of a half-drafted memo into a consumer chatbot, the upload of a spreadsheet nobody checked. That is genuinely the highest-volume risk, and killing it is a real win. What a block cannot do is stop a determined employee with a phone in their pocket. The personal device in every civil servant's coat is not covered by the Regeringskansliet's firewall, and the truly sensitive leak — the deliberate one, or the careless one at home — walks straight past it.
The internally developed platform is the tell. If the government believed ChatGPT and Claude were inherently unfit for official use, it would not be piping them into a controlled interface for selected staff. It believes they are unfit uncontrolled — which is a procurement position, not a security verdict. Fair enough, but it means Monday's headline overstates the transformation. The models are not leaving the building. They are being put on a leash, for some people, in some rooms.
The deeper risk is subtler: a single approved tool becomes a single point of failure and a single throat to choke. If Copilot has an outage, a vulnerability, or a contractual dispute, the ministries' AI capacity goes to zero overnight. Diversity of tools is itself a security property, and Stockholm just traded it away for auditability. Whether that trade was wise will be decided not by intranet memos but by the first incident.
Frequently asked questions
Is ChatGPT banned in Sweden?
No. There is no national ban and no regulator involved. Sweden's Government Offices are blocking ChatGPT and Claude on staff work computers only, as an employer IT-policy measure. Ordinary Swedes can use both tools freely.
What is the Regeringskansliet?
The Government Offices of Sweden — the collective administrative body serving the country's ministries, housed largely in the Rosenbad complex in central Stockholm. It is the employer in this story, not a regulator.
Why is Microsoft Copilot the only approved tool?
Because it is the tool the Government Offices have a formal enterprise agreement with, and whose security configuration they have reviewed and upgraded. The block is about contractual control and auditability, not about one model being technically safer than another — Copilot itself runs on OpenAI models.
Which other Swedish agencies have already done this?
According to Ekot's reporting, the Swedish Tax Agency and the National Board of Institutional Care have already blocked AI services on their systems, and the Enforcement Authority is considering following suit. The Government Offices had also previously restricted the Chinese chatbot DeepSeek.
How does this differ from Italy's 2023 ChatGPT ban?
Italy's Garante imposed a temporary nationwide block on ChatGPT in March 2023 over GDPR concerns; OpenAI made concessions and the ban was lifted that April. Sweden's move is an employer's network block affecting only government staff computers — no regulator, no citizen impact, and no company concession can reverse it.
Can staff still use ChatGPT or Claude at all?
On their office machines, no — from Monday the network block cuts direct access. But some personnel retain access to both models through an internally developed platform designed to interface with them safely, and nothing stops staff from using the consumer tools on personal devices outside work.
What have OpenAI and Anthropic said about the block?
Neither company has commented publicly on the block, per the coverage published today (Oct 23). OpenAI has not issued a statement, and Anthropic — whose Claude is being marketed aggressively as a workplace product — has likewise stayed silent.
Sources
Ekot / Sveriges Radio (original reporting, Oct 23, 2026), via relays: SVT, Daily Northern, AzerNews, Börsvärlden/Finwire, TechyNews; The Guardian (Kristersson AI-use reporting, Aug 2025); Reuters (Scaleway AMD supercomputer announcement, Oct 13, 2026; Scaleway press release); Hyperight and Computer Weekly (Sweden's February 2026 national AI strategy); The Hacker News / SecurityAffairs (Italy Garante reversal, April 2023). Internal links: our Scaleway AMD supercomputer in Sweden coverage, our Anthropic's push to turn Claude into a workplace staple, and our reporting on Spain's AI-agent breach notification.
Daniel Rao is a staff writer at TechPulse USA covering AI policy, regulation, and the geopolitics of compute. He previously covered enterprise software and has a weakness for government intranet memos.
Comments
Post a Comment