Summary
- Sweden's Government Offices restricted browser access to ChatGPT and Claude on work computers from 26 October.
- Microsoft Copilot remains approved, while the Intric platform provides access to the underlying AI models.
- The policy addresses unapproved AI use and information security; it is not a nationwide prohibition on generative AI.
Sweden’s Government Offices have restricted direct access to ChatGPT and Claude on official computers, directing employees towards approved artificial intelligence services as public institutions tighten control over the information processed by external platforms.
The browser restrictions took effect on 26 October, following a decision reported on 23 October. They prevent staff from opening the public versions of OpenAI’s ChatGPT and Anthropic’s Claude from government workstations, while retaining access to approved AI capabilities.
Employees can continue using Microsoft Copilot, which has undergone a security review, and the Government Offices have confirmed that ChatGPT and Claude models remain accessible through Intric, a managed AI assistant platform.
The decision addresses the use of AI services without appropriate contractual and technical arrangements. Sweden’s Government Offices are concerned that employees entering information into unapproved applications could expose official material to processing arrangements that have not been assessed or authorised.
Public interfaces give way to approved access
Generative AI applications have become common workplace tools because employees can use them through ordinary web browsers. Their capabilities include summarising documents, analysing information, drafting text and assisting with software development.
That accessibility creates a governance problem for employers handling sensitive material. An employee may submit information to an external service without the organisation having established the contractual permissions, retention conditions or security controls associated with its use.
Public AI services and business products built around similar models may operate under different terms. The treatment of submitted information can depend on the application, account configuration, contractual commitments and technical measures selected by the organisation.
Consequently, restricting an unapproved web interface does not necessarily mean an institution has rejected the underlying AI model. The same capabilities can be delivered through a managed environment with central administration, access controls and policies governing their use.
Sweden’s Government Offices have adopted that distinction. Staff are being directed towards applications the organisation has assessed rather than relying on individual judgement about the suitability of each external AI service.
The policy is preventive. Officials have raised concerns about the security consequences of unapproved use, but the restriction itself should not be interpreted as evidence that either OpenAI or Anthropic improperly handled a particular government dataset.
Intric provides an alternative route
Stockholm-based Intric secured a contract in July to provide an AI platform for the Government Offices. Its service allows organisations to build and manage assistants using their own information, with administrative controls over the environment.
The company describes a platform operating on Swedish infrastructure, with deployment options including managed services and installations in customer-controlled environments. The agreement is intended to provide AI capabilities without requiring employees to use unrestricted public interfaces.
Within that arrangement, Government Offices staff retain access to ChatGPT and Claude models. The precise availability of individual functions depends on the organisation’s implementation and permission arrangements rather than the capabilities of the public websites.
A managed platform can provide an organisation with greater visibility over which assistants exist, who can use them and which information sources they are permitted to access. It also enables common rules to be applied to services used by multiple departments.
Those controls remain dependent on configuration. Domestic hosting of an application does not, by itself, prove that every connected model, integration or supporting service processes data within the same jurisdiction.
Organisations must examine where information is processed, which parties have access and how data moves between the platform and any external systems. These distinctions are particularly important when public institutions handle personal information, confidential working documents and material concerning government decisions.
Microsoft Copilot remains permitted
The Government Offices have retained Microsoft Copilot as an approved AI service after reviewing and strengthening the security arrangements associated with its use.
Officials have not published detailed technical information about every control introduced during that review. It would therefore be premature to conclude that Copilot satisfies a particular set of protections beyond those the Government Offices have confirmed.
The decision illustrates how organisations can adopt different policies for applications offering related capabilities. Procurement terms, administrative tools, identity controls and data handling arrangements may all influence whether a service is suitable for official use.
Sweden’s Government Offices had previously restricted DeepSeek, while the Swedish Tax Agency and National Board of Institutional Care have introduced restrictions on AI services of their own. The Enforcement Authority has also been examining its position.
These are decisions taken by particular institutions. They do not establish a general national prohibition preventing all Swedish public employees from using ChatGPT, Claude or other generative AI products.
Unapproved AI becomes an operational risk
A review at the Swedish Transport Administration identified problematic use of several AI services, drawing attention to the extent to which employees can incorporate external tools into their work without a consistent organisational approval process.
Such activity is often described as shadow IT, where technology is adopted outside established procurement and management arrangements. Generative AI introduces additional concerns because employees may submit substantive business information directly to services that generate responses using external computing systems.
Restrictions on specific websites can reduce unapproved access, although similar functionality is increasingly embedded within productivity applications, development tools and other business software.
Maintaining effective control therefore requires organisations to understand where AI capabilities are available, which providers have been approved and what information employees may submit. Identity management, application controls and employee guidance address different parts of that requirement.
The Swedish policy preserves access to approved AI tools while removing direct access to two widely used public interfaces. Its implementation establishes a clearer boundary between applications available through ordinary browsers and those operating under the Government Offices’ authorised arrangements.
