Latest

Artificial Intelligence

Sweden bans ChatGPT and Claude at work: only Copilot

The Swedish Government has blocked ChatGPT and Claude on its officials' computers since 26 October and leaves Copilot as the only approved route.

Sweden bans ChatGPT and Claude at work: only Copilot
Sweden bans ChatGPT and Claude at work: only Copilot

Employees of the body that prepares the Swedish Government's decisions cannot open ChatGPT or Claude in the browser of their work computer. The block reached Regeringskansliet, the house that brings together the prime minister's office and ten ministries, with around 5,100 people on staff according to the official count.

, on Monday 26 October.

The reason is not that the models are worse, but that there is no contract setting out what happens to what is written inside them. Mikael Wettercrantz, the office's director of digitalization, warned on the intranet that using AI services without an agreement can have "consequences for our information security", and an internal audit at the Transport Administration had already described "problematic" uses of open platforms earlier this year.

The recommended alternative is Microsoft Copilot, which passed a security review and is covered by an agreement with the Swedish State. A limited number of employees keep access to ChatGPT and Claude through Intric, the assistant platform contracted in July, so the ban reaches the browser door and not the models. That distinction says more about the public procurement model than any press release.

A ban on the browser, not on the tool

The real scope is narrower than the headline suggests. The web interface of both services is no longer accessible on work computers, and the office has not clarified whether the block extends to personal phones or to programming interfaces. The decision became known on Friday 23 October, when theinternal announcement of the block

, was leaked. Afterwards, the press office declined to explain what changed in Copilot's configuration after the review, arguing that these are security matters, and Wettercrantz refused to give an interview. Two questions remain unanswered: what criteria make a tool approved and who audits that list months later.

Regeringskansliet is not a loose ministry. It is the machinery that prepares the Government's business and supports the ministers, with the prime minister's office and ten departments in the same administrative structure.

Government decisions, official inquiries and memoranda that may be classified are handled there. Of the nearly 4,900 people in active service, around 200 are political appointees and the rest keep their jobs regardless of who governs, so the ban mainly affects the permanent staff.

The report that lit the fuse

Before the ban there was an audit. The internal review by the Transport Administration detected "widespread and risky" use of around ten different artificial intelligence services and left a diagnosis on two fronts, that of the sensitivity of the information entered and that of the management of the responses generated.

That diagnosis describes the phenomenon jargon calls shadow IT. These are employees who get their work done with tools nobody has approved, without the IT department knowing what data leaves or which server it reaches.

With Copilot the administration stays within a contract; with ChatGPT in the browser, the text an official pastes about a file travels to a service that may keep it and reuse it to train its models. The ban did not come alone, because the office had already restricted DeepSeek along the same lines followed byUnited States federal systems

. The list of unapproved tools grows by way of data security, not by way of model quality. Body
Situation since 26 October Regeringskansliet (Government office)
Browser block on work computers Tax Agency (Skatteverket)
Block on open AI services National Board of Institutional Care (SiS)
Block on open AI services Enforcement Authority (Kronofogden)
Review under way before deciding Transport Administration (Trafikverket)

Prior internal audit with "problematic" use

Copilot wins by contract, not by model The easy reading is that Sweden has chosen one model and discarded two others, and the manufacturer's own documentation partly contradicts it. Copilot can run OpenAI models and also Anthropic ones, but in the European Union, the European Economic Area and the United Kingdom Anthropic models arrive disabled by default and remain outside the European data perimeter, according tothe instructions for administrators

. Someone has to enable them and assign users.

What separates Copilot from ChatGPT is not the weight of the model, but who answers if something breaks. There is a documented review, an agreement with the Swedish State and a configuration the house can touch, and that is what makes a tool approved. The US provider is still inside, with another contract and other conditions.

On 3 April this year Microsoft added a setting in its admin center so that Copilot uses Anthropic models in Microsoft 365 applications within the EU, the EEA and the United Kingdom. The door exists, but it is closed by default and an administrator decides who goes through it.

The back door is called Intric The exception portrays the Swedish public procurement model better than the ban does. The platform that keeps access to the two banned models is Intric, a Stockholm company that won the contract on 8 July for twelve months, extendable by another twelve, and that runs the assistants onSwedish infrastructure

, in a local cloud, on client servers or even isolated from the internet. The technical detail matters, because the platform is model-agnostic and allows choosing between ChatGPT, Claude, Gemini or open weights depending on the task. It fits the debate on closed and open models

that the labs themselves are waging. An official can still ask Claude what a memorandum says, as long as the query goes through a system the State controls and hosts.

That arrangement has a review date and an implicit incentive: if the platform cannot withstand the daily use of thousands of officials, the approved alternative is still Copilot and the State returns to the single provider.

The bill for depending on a single provider

The underlying problem remains, and it is not technical. A public body that reduces its perimeter to a single provider swaps a diffuse risk for a concrete dependency, and the office has not published what security adjustments it made to Copilot or by what criteria it evaluated the alternatives. Transparency is precisely what allows the ban to be an administrative decision and not an act of faith.

The political detail adds spice to the scene. The prime minister, Ulf Kristersson, once said he used AI models "quite often" in his work, and now his officials have to go through Copilot or an internal platform to do the same. The question is no longer whether AI enters the administration, but who signs the contract that lets it in.

Who decides which tools the State uses

The Swedish sequence will be repeated in more capitals, and not because of a regulatory fashion but because of provider arithmetic. More and more public work goes through conversational assistants, and each one is a transfer of data to a foreign jurisdiction. The Tax Agency and the National Board of Institutional Care have already closed the door, and the Enforcement Authority is studying doing so.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading