Sweden bans ChatGPT and Claude for government use while approving Microsoft Copilot

The move to enterprise AI

Sweden’s central government is tightening its control over how civil servants interact with artificial intelligence. The Government Offices have officially blocked the use of consumer AI services, specifically naming ChatGPT and Anthropic’s Claude, on staff computers. This decision aims to prevent the unauthorized entry of sensitive government data into public-facing systems.

The primary reason for this restriction involves the legal and contractual differences between consumer products and enterprise tools. Mikael Wettercrantz, the director of digitalisation for the Government Offices, explained that using consumer services without a formal contract can have consequences for information security. He noted that these tools often route queries to external company servers where the data may be subject to U.S. law.

By contrast, the government has designated Microsoft Copilot as the only approved assistant for civil servants. This approval stems from a vetted enterprise agreement that sets specific limits on what a provider can do with government inputs. The office reviewed Copilot and tightened its security settings to ensure it meets these necessary standards before clearing it for use.

Security risks and data flow

The government's concern centers on the lack of clarity regarding how AI-generated outputs are handled and where data is stored. When employees use consumer AI tools, they may inadvertently share sensitive information with a company that does not have a binding agreement to protect that specific data. This creates a significant risk for national security and privacy.

The distinction between a public tool and a private contract is central to the Swedish government's strategy. A vetted agreement provides a legal shield that limits a provider's ability to use or share government data. Consumer tools generally do not offer these same protections, making them unsuitable for handling any information that requires high levels of confidentiality.

Furthermore, the government is concerned about the

A growing trend of European scrutiny

Sweden's decision is part of a broader wave of actions across Europe regarding AI safety and data privacy. The European Parliament took a similar stance in February by disabling AI features on all parliamentary devices. They concluded that they could not guarantee the security of the data being processed by these systems.

This trend of scrutiny is not new and can be traced back at least to a ban on ChatGPT in Italy in 2023. These repeated actions across different nations suggest a growing consensus that public institutions require more than just product claims to feel safe using AI.

The complexity of approved tools

The distinction between what is banned and what is approved is not always straightforward. For example, Microsoft added Claude models into Copilot in early 2026. Additionally, Anthropic’s models in the EU are set to off by default and require an administrator to enable them manually.

Despite these technical nuances, the Swedish government has not confirmed whether its specific Copilot instance has Claude switched on. This lack of clarity shows that even within "approved" enterprise tools, the specific configurations and active models can vary significantly between different organizations.

Implications for data sovereignty

For the crypto and Web3 community, these government restrictions serve as a signal that the custody and decentralized control of data are becoming increasingly important. As institutions move away from public tools, there is a growing demand for verifiable data-handling rules and private inference methods.

This shift suggests that self-hosted models and private deployments may become more attractive to public-sector customers. These technologies promise tighter custody of inputs and outputs, addressing the legal exposure created by overseas servers and foreign laws.

Internal access and hidden rules

While the government has blocked direct access to consumer AI, the situation is not a total shutdown for all staff. Some employees will still maintain the ability to reach ChatGPT and Claude through a custom internal AI platform developed by the Government Offices. However, the authorities have not yet shared details regarding how many people have access to this specific system or what safety guardrails are in place.

This internal workaround creates a complex environment where different types of access coexist within the same agency. Because the specific rules for this platform remain undisclosed, it is unclear how the government intends to balance the utility of these models with the security concerns that led to the initial ban. This highlights the ongoing challenge of balancing innovation with strict data protection protocols.

Data custody and decentralized trends

The move toward vetted enterprise contracts and private deployments reflects a broader shift in how institutions view data sovereignty. For the crypto and Web3 community, these government actions serve as a clear signal that the custody and decentralized control of data are becoming primary concerns for public-sector customers. As legal exposure from overseas servers grows, the demand for verifiable data-handling becomes a critical requirement.

This trend suggests that technologies offering private inference and self-hosted models may see increased interest from government bodies. By prioritizing systems that promise tighter custody of both inputs and outputs, public institutions are looking for ways to mitigate the risks posed by foreign laws. This shift underscores a growing preference for technical guarantees over simple product claims in the AI landscape.