On Monday morning, staff at Sweden’s Government Offices will open their laptops and find two familiar buttons gone. Sweden blocks ChatGPT and Claude on work computers starting October 26, and the official reason is stranger than any ban story you’ve read. It’s not about what the models said, it’s about who signed the contract.
Sweden blocks ChatGPT and Claude: what actually happened
The Government Offices (Regeringskansliet) is the civil service behind the Swedish state: the Prime Minister’s Office, the ministries, the administrative core. From October 26, the browser versions of ChatGPT and Claude stop loading on their work computers. Staff get pointed to Microsoft Copilot instead, now the only chatbot approved for direct use.
The news broke through Sweden’s public broadcaster, and the Government Offices confirmed it to SVT. One thing to get straight: this is not a national ban. Private users and companies in Sweden can keep doing whatever they were doing. It covers government work devices only.
Why Copilot survived and the others did not
| Tool | Status from October 26 | Why |
|---|---|---|
| ChatGPT | Blocked | No contract, no security review |
| Claude | Blocked | No contract, no security review |
| Microsoft Copilot | Approved | Reviewed, hardened, contracted |
| Intric (internal platform) | Allowed for some staff | Controlled access to the same models |
Mikael Wettercrantz, the office’s director of digitalisation, wrote on the staff intranet that using AI services without a formal contract can have “consequences for our information security.” He also flagged “several examples of problematic use of open AI services,” covering both how sensitive the pasted information was and how staff handled the AI’s answers afterwards.
Copilot passed because the office reviewed it, raised its security settings, and signed a contract. The models were never really the issue. Data handling was. On a consumer plan, the provider’s standard terms decide where your prompts and files live and whether they train anything. An organization contract sets those rules in advance, in writing, with someone accountable when things go wrong. In other words, Sweden blocks ChatGPT and Claude for one deeply boring reason: procurement.
The awkward part nobody mentions
Here’s the detail that makes the ban look less clean than the headline: Microsoft folded Claude models into Copilot in early 2026. Inside the EU, Anthropic’s models sit in Copilot switched off by default, and an administrator has to flip them on. Sweden’s Government Offices has not said whether its own Copilot has them enabled. So the “banned” AI might literally be running inside the “approved” one, behind a toggle nobody will confirm.
There’s an exception too. Some staff keep using ChatGPT and Claude through Intric, an internal AI platform the Government Offices built itself. Same models, controlled pipe. That exception tells you the real policy: the models were never the problem. Uncontrolled data paths were.
Not everyone in Sweden agrees on the risk
Here’s where it gets interesting: other Swedish agencies ran the same tools and reached different conclusions. The Swedish Post and Telecom Authority piloted both Copilot and ChatGPT under strict information-handling guidelines. Kammarkollegiet, the legal oversight agency, tested both before rolling out digital assistants to all its employees. The National Institute of Economic Research even rated the information-security risk of tools like Copilot as very low.
So while Sweden blocks ChatGPT at the Government Offices, DeepSeek was already restricted there, the Tax Agency and the National Board of Institutional Care have similar blocks in place, and the Enforcement Authority is still deciding. One country, one government, several different answers. Expect the same inconsistency inside any big organization, including yours.
Why this matters if you don’t work in government
You’re not Swedish civil service, so why care? Because the pattern is spreading fast. The European Parliament went first in February, disabling AI features on every work device (ChatGPT, Claude, and Copilot included) after deciding it couldn’t guarantee how those tools handled data. Sweden’s Tax Agency and the National Board of Institutional Care already block similar services, and the Enforcement Authority is reviewing its own rules.
If your employer touches contracts, health data, client files, or anything with an NDA attached, expect an AI policy memo eventually. Better to be ahead of it. The Swedish setup is a decent preview of what those policies look like in practice: one approved tool backed by a contract, everything else off the work laptop, and a controlled escape hatch for people who genuinely need more.
A 3-minute check before you paste anything
1. Look at what you’re about to paste. If a colleague would be annoyed to see it screenshotted somewhere public, it doesn’t belong in a consumer chatbot. Picture the quarterly forecast with client names on it: paste it into a free chatbot and you’ve just shared it with a third party, on terms you never read.
2. Check your plan. Paid tiers from OpenAI and Anthropic generally exclude your data from model training, but a signed company contract with real data rules is the version your legal team will love.
3. Ask whether your workplace already has an approved tool. Use it for anything work-adjacent, even if the consumer version feels nicer. Approval is about the contract, not the chatbot.
4. Keep the split clean. Work device, approved tool. Personal phone, personal account. Mixing the two is exactly what tripped up the Swedish staff.
Your takeaway
Sweden blocks ChatGPT and Claude at work not because the tools got worse, but because nobody signed for them. That’s the whole story in one line, and it’s the same line your IT department will use eventually. Check what your workplace allows this week, before a memo writes the rules for you.
If you want the privacy side in more depth, we’ve covered how ChatGPT ad targeting uses your prompts, the AI agent Slack leak that should worry you, and how to stop Google from demanding your phone number. All three pair nicely with this one.