Sweden blocks ChatGPT and Claude on government computers

A targeted workplace restriction

Sweden’s Government Offices officially blocked browser-based access to OpenAI’s ChatGPT and Anthropic’s Claude on employee computers on Monday, October 26, 2026. As reported by shattered.io, the move was a specific response to security concerns regarding how staff interact with these services.

It is important to note that this is a workplace access rule rather than a national ban on the products themselves. The restriction specifically targets the public web versions that workers can open in a browser tab to paste text without any formal approval or procurement process.

The scope of this decision is limited to the Government Offices, which serves the Swedish government. Other entities, such as municipalities and regional bodies, run their own IT policies and have not been confirmed to follow this specific block.

The missing data agreement

The primary reason for the block is the absence of a formal agreement governing data protection and processing. Officials are concerned about employees using AI services that do not have a clear paper trail defining who processes the data and for what purpose.

Without a contract, a government agency cannot prove to auditors what happens to a draft memo or a piece of sensitive text once it is pasted into a third-party chat box. This distinction explains why Microsoft Copilot survived the cut, as it sits within an existing Microsoft environment and established agreements.

Mikael Wettercrantz, the digitalization director at the Government Offices, warned that using AI without such agreements can carry consequences for information security. The focus is on the flow of data rather than the capabilities of the models themselves.

Addressing problematic use

The government identified two specific types of problematic use that prompted this review. The first involves the sensitivity of information being entered into prompts, where staff might paste entire email threads or draft legislation into a free consumer account.

The second concern involves how employees handle the AI-generated responses. If a staffer treats a generated answer as finished work without verifying it, errors or biases could carry the weight of the institution. The government also flagged risks regarding how generated text might behave when pasted into other systems.

By blocking the easiest route to these tools, the government aims to reduce 'shadow AI'—the practice of employees using unsanctioned tools to save time. They are attempting to move users toward tools where the organization can log requests and enforce security standards.

A two-tier access system

While the public web versions are blocked, the government is not completely rejecting these models. Some employees will retain access to ChatGPT and Claude through a special AI-assistant platform developed or provided specifically for the Government Offices.

This managed platform acts as a middle layer between the employee and the model. It allows the agency to apply logging, restrict access, and route requests under negotiated terms that a standard browser tab cannot provide. This creates a two-tier system for AI usage within the office.

Most staff will use Copilot as their primary tool, while a subset of employees with defined needs will use the managed platform. This structure allows the government to balance the need for AI productivity with the requirement for strict information security.

Precedent for other agencies

This decision serves as a template for other government agencies that want to manage AI risk without engaging in a political fight over specific vendors. Because the objection is about contracts and controls rather than a claim of wrongdoing by OpenAI or Anthropic, other bodies can copy the logic.

The move highlights that approval follows the presence of a signed data-processing agreement rather than the strength of a model's performance. A different vendor with a signed agreement could have occupied the same approved slot as Microsoft.

For the AI industry, the impact is less about immediate revenue and more about procurement behavior. It signals that for government work, the ability to prove data sovereignty is the most critical requirement for any AI product.